Sherry: Hardware-Efficient 1.25-Bit Ternary Quantization via Fine-grained Sparsification Hong Huang Decheng Wu Qiangqiang Hu Guanghua Yu Jinhai Yang, Jianchen Zhu Xue Liu Dapeng Wu Abstract The deployment of Large Language Models (LLMs) on resource-constrained edge devices is increasingly hindered by prohibitive memory and computational requirements. While ternary quantization offers a compelling solution by reducing weights to { − 1 , 0 , + 1 } \{-1,0,+1\} , current implementations suffer from a fundamental misalignment with commodity hardware. Most existing methods must choose between 2-bit aligned packing, which incurs significant bit wastage, or 1.67-bit irregular packing, which degrades inference speed. To resolve this tension, we propose Sherry , a hardware-efficient ternary quantization framework. Sherry introduces a 3:4 fine-grained sparsity that achieves a regularized 1.25-bit width by packing blocks of four weights into five bits, restoring power-of-two alignment. Furthermore, we identify weight trapping issue in sparse ternary training, which leads to representational collapse. To address this, Sherry introduces Arenas , an annealing residual synapse mechanism that maintains representational diversity during training. Empirical evaluations on LLaMA-3.2 across five benchmarks demonstrate that Sherry matches state-of-the-art ternary performance while significantly reducing model size. Notably, on an Intel i7-14700HX CPU, our 1B model achieves zero accuracy loss compared to SOTA baselines while providing 25% bit savings and 10% speed up. The code is available at https://github.com/Tencent/AngelSlim . 1 Introduction Large language models (LLMs) have demonstrated remarkable capabilities across massive applications Wu et al. 2023 ; Floridi and Chiriatti 2020 . However, the increasing concerns regarding data privacy, the necessity for offline functionality, the prohibitive latency, and the cost of cloud-based inference ( Yao et al. 2024 ; Liagkou et al. 2024 ) have driven a pressing need to deploy LLMs on resource-constrained edge devices. Figure 1: Comparison of different packing strategies for ternary quantization in efficiency. Weight quantization ( Hubara et al. 2018 ; Liu et al. 2023 ) is a practical technique for enabling on-device deployment by reducing model footprints and computation through lower-precision numerical representations. Nevertheless, most existing quantization methods ( Dettmers et al. 2024 ; Lin et al. 2023b ; Frantar et al. 2022 ) are optimized for server-grade GPUs that support complex hardware primitives, such as mixed-precision multiplication. Thus, these methods are often unsuitable for heterogeneous edge and mobile hardware, which prioritize simplified, widely supported operations that avoid specialized hardware dependencies. Figure 2: (Left) 2-bit strategy packs each weight into 2 bits to maintain alignment, resulting in large bit wastage. (Middle) 1.67-bit strategy packs 3 weights into 5 bits, introducing SIMD-unfriendly 3-way patterns, leading to slow speed. (Right) Our Sherry enforces a 3:4 sparsity and packs 4 weights into 5 bits, introducing SIMD-friendly 4-way patterns, achieving a small 1.25-bit width and faster inference speed. More details are shown in Fig 9 . Ternary quantization ( Li et al. 2016 ; Liu and Liu 2023 ) provides a compelling paradigm for edge-based LLM inference by constraining weights to the set \{-1,0,+1\} . Supported by the Lookup Table (LUT)-based engines Huang et al. 2025c ; Wei et al. 2025 ; Wang et al. 2025b ; Nie et al. 2025 , ternary quantization transforms costly floating-point multiplications into hardware-efficient additions, as shown in Fig 9 . This intrinsic compatibility makes ternary quantization a promising solution to bridge the gap between high-performance LLMs and the strict resource constraints of edge devices. However, existing ternary quantization methods ( Ma et al. 2025 ; Wang et al. 2025a ; Huang et al. 2025a ) suffer from practical inefficiencies due to the misalignment between non-standard ternary bit-widths and standard hardware architectures. Current implementations generally resort to two suboptimal strategies for LUT-based execution: (1) 2-bit strategy that packs each weight into 2 bits ( Wei et al. 2025 ) , as shown in Fig. 2 (left) , offering no memory savings over standard INT2 quantization; or (2) 1.67-bit strategy that packs three weights into five bits ( Wang et al. 2025b ) , as shown in Fig. 2 (middle) . While the 1.67-bit strategy offers improved density, its 3-way grouping is fundamentally incompatible with the power-of-two vector lanes of modern Single Instruction Multiple Data (SIMD) units, frequently leading to a slower inference speed compared to the 2-bit strategy. Consequently, existing ternary methods are forced to trade off between bit width and inference speed , preventing them from achieving the full theoretical benefits of ternary weights. To address these limitations, we propose Sherry , a novel S parse h ardware- e fficient t er na r y quantization framework that achieves 1.25-bit width while maintaining superior inference speed. Our key insight is that the inherent sparsity of ternary models can be strategically structured to reconcile the tension between storage density and computational regularity. Specifically, Sherry enforces an optimal 3:4 fine-grained sparsity constraint: within every contiguous block of four weights, exactly three are quantized to non-zero values ( \pm 1 ), and one is fixed to zero. This structured constraint enables an optimal packing strategy where each 4-weight block is stored in a compact 5-bit representation. Crucially, this block-based approach restores the power-of-two alignment required by modern SIMD units, allowing for parallel processing in regularized hardware operations. Furthermore, directly integrating 3:4 sparsity into ternary training often results in performance degradation. We attribute this to weight trapping : weights accumulate in localized regions due to gradient homogenization , leading to representational collapse, as shown in Fig. 3 and 11 . To address this, Sherry introduces Arenas 1 1 1 ”Arenas” is also the name of the sandy soil used to grow grapes for sweet Sherry wine. , an A nnealing re sidual sy na p s e module. By injecting heterogeneous gradients during the training phase, this module breaks the gradient homogenization and maintains an expressive and diverse weight distribution for Sherry. We evaluate the efficacy and efficiency of Sherry across five standard benchmarks using the LLaMA-3.2 ( Touvron et al. 2023 ) model family. Our empirical results demonstrate that Sherry matches performance with state-of-the-art (SOTA) ternary quantization baselines while necessitating a substantially lower bit width. Notably, on an Intel i7-14700HX CPU, Sherry achieves a 16% reduction in model size and a 10\% inference speed up compared to SOTA 1B ternary LLM, with zero accuracy degradation. These findings validate that Sherry provides a promising hardware-efficient solution for deploying LLMs on resource-constrained platforms. 2 Background and Challenge 2.1 Ternary Quantization Ternary quantization is an extreme weight compression paradigm that constrains model parameters to the discrete set \{-1,0,+1\} . Taking per-channel quantization as an example, for a full-precision weight matrix W\in\mathbb{R}^{d_{in}\times d_{out}} , where d_{in} is the number of input channels and d_{out} is the number of output channels, the general ternary quantization function Q(\cdot) is defined as: Q(W)=T\alpha,\quad T_{i,j}=\begin{cases}+1,&\text{if }W_{i,j}>\Delta_{j};\\ 0,&\text{if }|W_{i,j}|\leq\Delta_{j};\\ -1,&\text{if }W_{i,j}<-\Delta_{j},\end{cases} (1) where T\in\{-1,0,+1\}^{d_{in}\times d_{out}} is the ternary weight matrix, \alpha\in\mathbb{R}^{d_{out}} represents the scaling factors 2 2 2 In this paper, we denote the multiplication between matrix T and vector \alpha as element-wise, i.e. , [T\alpha]_{i,j}=T_{i,j}\alpha_{j} . , and \Delta\in\mathbb{R}^{d_{out}} denotes the quantization thresholds. A substantial body of research has investigated the optimal determination of \alpha and \Delta , which are discussed in detail in Sec. E . To eliminate weight-activation multiplications during inference, Lookup Table (LUT)-based engines Wang et al. 2025b ; Wei et al. 2025 have been developed, demonstrating superior efficiency over traditional multiplication-based engines. As illustrated in Fig. 9 , the engine segments the inputs and pre-computes a localized lookup table. For each segment, the corresponding ternary weight index is used to retrieve pre-computed results from the table. This paradigm provides a promising pathway for ternary LLMs by fully replacing complex floating-point multiplications with highly efficient additions and memory lookups. 2.2 Quantization-Aware Training (QAT) Due to the aggressive nature of ternary compression, QAT is essential to recover model fidelity. In the forward pass, full-precision weights W are dynamically quantized via Q(\cdot) in Eq. 20 . During the backward pass, since Q(\cdot) is non-differentiable, we employ the Straight-Through Estimator (STE) ( Zhu et al. 2016 ; Chen et al. 2024b ) to approximate gradients. For an input matrix X\in\mathbb{R}^{d_{t}\times d_{in}} , where d_{t} is the number of tokens, the forward pass and gradients for loss L are: Y=XT\alpha,\quad\frac{\partial L}{\partial W}\approx X^{\top}\frac{\partial L}{\partial Y}. (2) After training, the weights W are discarded, leaving the ternary weights T and scaling factors \alpha . 2.3 Challenge Although ternary quantization has a theoretical lower bound of 1.58 bits ( \log_{2}3 ), contemporary implementations encounter significant architectural friction when deployed on commodity hardware. Most current methods resort to two suboptimal strategies: (1) 2-bit Strategy: This approach pads each ternary weight into 2 bits ( Wei et al. 2025 ) , as shown in Fig. 2 (left) . While it preserves computational regularity and aligns with SIMD vector lanes, it effectively nullifies the storage advantages of the ternary set compared to 2-bit integer quantization. (2) 1.67-bit Strategy: This scheme packs three weights into 5-bit blocks ( Wang et al. 2025b ) , as shown in Fig. 2 (middle) . However, this introduces severe arithmetic inefficiencies because modern hardware accelerators are optimized for 2^{n} operand groupings. In practice, this misalignment often results in slower inference speeds than the 2-bit strategy. Alternative strategies, such as packing 5 weights into 8 bits ( 1.6 -bit), exponentially increase the size of the lookup tables, rendering them unsuitable for LUT-based edge engines. Consequently, existing ternary quantization remains trapped in a trade-off between memory footprint and inference speed, preventing it from realizing its full theoretical benefits in practical edge deployments. Figure 3: Comparison of weight distributions for LLaMA-1B under different quantization schemes. (Left) Naive 3:4 sparse ternary training exhibits weight trapping , where values collapse into a binary-like polarization, leading to a significant accuracy gap compared to dense models. (Right) Our Sherry utilizes the Arenas module to achieve a trap-free distribution, bridging the performance gap to dense ternary models. 3 Proposed Sherry 3.1 3:4 Sparse Ternary Quantization The primary objective of Sherry is to achieve substantial bit savings without compromising inference throughput. Building on the observation that ternary models exhibit inherent sparsity Liu and Liu 2023 , we formalize this characteristic by adopting an N:M structured sparsity constraint. This enforces a rigid pruning pattern where exactly N non-zero elements are permitted within every contiguous block of M weights. While traditional N:M sparsity (e.g., the 2:4 pattern) is typically coupled with specific GPU-vendor kernels ( Zhou et al. 2021 ) , Sherry leverages a multiplication-free LUT engine to decouple the sparsity constraint from specialized hardware primitives. This architectural freedom allows us to explore a 3:4 structured sparsity pattern, which achieves a highly efficient 1.25-bit width. The optimization goal of Sherry is to minimize the L_{2} reconstruction error between the full-precision weights W and the sparse ternary representation T\alpha , subject to the 3:4 sparsity constraint. Let W_{b:b+3,j}=(W_{b,j},W_{b+1,j},W_{b+2,j},W_{b+3,j}) denote a contiguous vector. The objective is formulated as 3 3 3 In this paper, we denote W_{:,j}\in\mathbb{R}^{d_{in}} as the j -th column o matrix W\in\mathbb{R}^{d_{in}\times d_{out}} , i.e., W_{:,j}=(W_{1,j},\dots W_{d_{in},j}) : \displaystyle\min_{T,\alpha} \displaystyle\sum_{j=1}^{d_{out}}\|W_{:,j}-T_{:,j}\alpha_{j}\|_{2}^{2} (3) \displaystyle\text{s.t. } \displaystyle T_{i,j}\in\{-1,0,+1\}, \displaystyle\forall b\in\{1,5,\dots,d_{in}-4\}:\|T_{b:b+3,j}\|_{0}=3, where \|\cdot\|_{0}=3 enforces the 3:4 sparsity by ensuring exactly three non-zero values per block. The optimal solution to the objective in Eq. 3 is obtained via a greedy Sparse-AbsMean strategy. For each block W_{b:b+3,j} , we prune the element with the smallest absolute magnitude and assign ternary values to the remaining three. Formally, for each block i\in[b,b+3] , the optimal ternary element T_{i,j} is given by: T^{*}_{i,j}=\begin{cases}0,&\text{if }i=\arg\min|W_{i,j}|;\\ \text{sign}(W_{i,j}),&\text{otherwise}.\end{cases} (4) Given this optimal ternary matrix T , the optimal scaling factor \alpha_{j} for the j -th output channel is calculated as the mean absolute value of the non-pruned weights: \alpha^{*}_{j}=\frac{4}{3d_{in}}\sum_{i\in\mathcal{S}_{j}}|W_{i,j}|, (5) where \mathcal{S}_{j}=\{i\mid T_{i,j}\neq 0\} denotes the set of active (non-zero) indices in the j -th column. The proof of the optimal is presented in Sec. D . Although rarely used in floating-point contexts, 3:4 sparsity represents an ideal "sweet spot" for ternary quantization across four dimensions: Figure 4: The effective ranks (ER) of the gradients during training. Both Binary and 3:4 sparse ternary training have relatively low ER due to gradient homogenization. Figure 5: The overview of the Arenas module with QAT. The Arenas module injects the heterogeneous gradients through a residual synapse with an annealing gate. (1) SIMD-friendly Alignment: The choice of M=4 ensures power-of-two alignment with activation segments, which is critical for SIMD vector lane loading. This eliminates the complex bit-shuffling overhead characteristic of 1.67-bit (3-way) packing schemes. (2) Safe Sparsity Margin: Prior research Zhu et al. 2016 indicates that ternary quantization performance undergoes severe degradation when the sparsity ratio exceeds 50%. 3:4 structured scheme maintains a 25% sparsity level within the safe margin required to preserve model expressive capacity. (3) Optimal Bit-State Utilization: For M=4 , enforcing N=3 non-zero elements yields exactly \binom{4}{3}\times 2^{3}=32 unique permutations. This mathematically saturates a 5-bit index ( 2^{5}=32 ), ensuring maximum bit-state utilization without bit wastage. (4) Compatibility with LUT-Inference: The 3:4 pattern is natively compatible with the LUT-based ternary inference engine, as shown in Fig 9 . Given the 128-bit constraints of standard SIMD register instructions (e.g., AVX2 vpshufb), the maximum capacity for a single-instruction lookup table is 16 bytes. By utilizing mirror-symmetry of ternary states Ma et al. 2025 ; Wei et al. 2025 , Sherry splits the 5-bit representation into 1 sign bit and 4 index bits, perfectly fitting within search limits. In summary, the 3:4 format represents an intersection of hardware regularity and representational fidelity. We prove that the 3:4 format is an optimal for the LUT-based engine in Sec. C . 3.2 Arenas: Annealing Residual Synapse Despite the hardware and bit-utilization advantages of 3:4 structured sparsity, its direct application in ternary quantization often triggers significant performance degradation compared to the dense model. We identify the root cause as weight trapping . As illustrated in Fig. 3 , under the hard 3:4 pruning constraint, weights tend to polarize toward specific values, resulting in a distribution that mimics binary quantization. This collapse prevents the model from exploiting the expressive capacity of the ternary set, essentially trapping it in a suboptimal binary-like state in Fig. 10 (Top Right) . We find that this stagnation is driven by Gradient Homogenization . In a 3:4 sparsity configuration, the zeros are distributed uniformly within T , causing the sparse matrix to behave similarly to a dense binary matrix. This uniform distribution mimics the properties of a Hadamard transform Huang et al. 2024b ; Ashkboos et al. 2024 , which flattens the signal. Consequently, the downstream gradients w.r.t. the activations X become increasingly homogenized: \frac{\partial L}{\partial X}=\frac{\partial L}{\partial Y}(T\alpha)^{\top}. (6) Therefore, the gradients passed to preceding layers become undifferentiated. This forces the majority of weights to exhibit similar behavior during training, significantly decreasing the representational diversity. In Fig. 5 (left) , the Effective Rank (ER) Roy and Vetterli 2007 for 3:4 sparse training confirms our analysis. Specifically, the 3:4 sparse regime exhibits a low ER ( ER<750 ), a level of spectral collapse comparable to that of binary quantization, despite the gradient matrix having a total dimensionality of 4096. This indicates a significant loss of learning degrees of freedom in naive 3:4 ternary sparse training (see Appendix F for more details). To restore representational diversity, we propose Arenas, an Annealing Residual Synapse mechanism that re-couples latent weight magnitudes to the loss objective via a continuous bypass. During the training phase, the output of a ternary linear layer is augmented with a decaying full-precision residual synapse: Y=XT\alpha+\lambda_{t}XW, (7) where \lambda_{t} is a scheduling coefficient that anneals to zero by the conclusion of training. The inclusion of the latent matrix W in the forward pass fundamentally alters the gradient dynamics. The gradient with respect to the latent activation X becomes: \frac{\partial L}{\partial X}=\frac{\partial L}{\partial Y}(T\alpha+\lambda_{t}W)^{\top}. (8) By incorporating the continuous values of W into the backward path, Arenas injects heterogeneous information back into \frac{\partial L}{\partial X} , breaking the homogenization effect induced by the 3:4 structure. This allows earlier layers to receive specialized updates, providing the "energy" required for weights to escape trapped binary-like states. As \lambda_{t}\to 0 , the residual vanishes, leaving a pure 3:4 sparse ternary model for inference with zero additional overhead. Size Method Bit-width ARC-e ARC-c HelS PIQA WinG Average 1B BF16 16 0.654 0.313 0.477 0.742 0.603 0.558 LSQ 1.67 0.376 0.177 0.258 0.574 0.506 0.378 SEQ 1.67 0.421 0.180 0.273 0.604 0.510 0.398 DLT 1.67 0.424 0.174 0.256 0.563 0.513 0.386 TWN 1.67 0.407 0.220 0.284 0.601 0.492 0.401 AbsMedian 1.67 0.567 0.251 0.339 0.674 0.533 0.473 AbsMean 1.67 0.603 0.259 0.360 0.683 0.541 0.489 Tequila 1.67 0.645 0.305 0.391 0.710 0.542 0.519 Sherry 1.25 0.647 0.309 0.388 0.699 0.550 0.519 3B BF16 16 0.745 0.422 0.552 0.768 0.691 0.636 LSQ 1.67 0.431 0.200 0.294 0.599 0.522 0.409 SEQ 1.67 0.498 0.231 0.303 0.645 0.529 0.441 DLT 1.67 0.361 0.161 0.260 0.572 0.496 0.370 TWN 1.67 0.692 0.351 0.462 0.734 0.586 0.565 AbsMedian 1.67 0.636 0.299 0.406 0.713 0.558 0.522 AbsMean 1.67 0.672 0.329 0.439 0.735 0.582 0.551 Tequila 1.67 0.702 0.346 0.464 0.739 0.627 0.576 Sherry 1.25 0.688 0.364 0.452 0.736 0.593 0.567 Table 1: Comparison of Sherry with different ternary quantization methods. The Arenas provides three critical advantages for 3:4 sparse ternary training: (1) Variance Injection and Singularity Breaking: By re-introducing the continuous matrix W , Arenas prevents the gradients \frac{\partial L}{\partial X} from collapsing into a homogenized, low-rank state, allowing earlier layers to receive specialized updates (in Fig. 5 ). (2) Adaptive Error Compensation: During training, the residual term \lambda_{t}XW naturally absorbs the quantization noise and pruning error introduced by the 3:4 ternary constraint. This allows the network to maintain a high-precision internal representation while the sparse ternary weights ( T\alpha ) gradually learn to capture the most salient components of the signal, leading to better optimization. (3) Zero-Overhead Inference: Because \lambda_{t} anneals to zero, the auxiliary path is completely removed post-training, without any inference overheads. 4 Evaluation To validate the efficacy of Sherry, we conduct experiments evaluating its performance against ternary quantization baselines. All results are averaged over three independent runs with random seeds. In all tables, the best and second-best results are highlighted in purple and blue ; half-precision (BF16) results are shown in gray for reference. 4.1 Experimental Setup We provide an overview of our experimental configuration below; further implementation details are available in Appendix G . Datasets, Models, and Evaluation: We utilize LLaMA-3.2-1B and LLaMA-3.2-3B ( Touvron et al. 2023 ) as our base architectures. Unless otherwise specified, we employ group-wise quantization with group size 128. For quantization-aware training, we utilize 10B tokens sampled from the UltraFineWeb dataset ( Wang et al. 2025c ) . Following established ternary quantization benchmarks ( Liu et al. 2025 ; Chen et al. 2024b ; Ma et al. 2025 ) , we evaluate performance using lm-evaluation-harness ( Gao et al. 2024 ) across five zero-shot tasks: PIQA ( Bisk et al. 2020 ) , ARC-Easy (ARC-e), ARC-Challenge (ARC-c) ( Clark et al. 2018 ) , HellaSwag (HelS) ( Zellers et al. 2019 ) and WinoGrande (WinG) ( Sakaguchi et al. 2021 ) . In our results, we report a bit-width of 1.67 for existing ternary quantization baselines rather than the theoretical 1.58 , as 1.67 reflects the actual bit-density achieved by current practical packing strategies. Ternary Quantization Baselines: We compare Sherry against two categories of quantization methods: (1) Static methods , including TWN Li et al. 2016 , Tequila used in TequilaLLM Huang et al. 2025a , and the AbsMedian/AbsMean strategies utilized in BitNet ( Ma et al. 2025 ; Wang et al. 2023 ) and Spectra ( Kaushal et al. 2025 ) ; and (2) Learnable methods , such as DLT ( Chen et al. 2024b ) used in TernaryLLM, LSQ ( Esser et al. 2019 ) , and the SEQ strategy used in ParetoQ ( Liu et al. 2025 ) . Implementation Details: Training is conducted on 32 NVIDIA GPUs, while inference throughput is evaluated on an Intel 8263C CPU to verify edge-efficiency. We quantize all linear layers within the Transformer architecture using a sequence length of 1024. The learning rate is fixed at 10^{-4} . The annealing coefficient \lambda_{t} follows a cosine-decay schedule with warmup by default, as shown in Fig. 8 . Model Size Bit-width ARC-e ARC-c HelS PIQA WinG Average LLaMA3.2 1B 16 0.654 0.313 0.477 0.742 0.603 0.558 TernaryLLM ∗ 1B 1.67 0.424 0.174 0.256 0.563 0.513 0.386 ParetoQ ∗ 1B 1.67 0.421 0.180 0.273 0.604 0.510 0.398 LLM-QAT 1B 1.67 0.360 0.262 0.313 0.551 0.496 0.397 BitNet 1.3B 1.67 0.549 0.242 0.377 0.688 0.558 0.483 Spectra 1.1B 1.67 0.563 0.246 0.388 0.693 0.555 0.489 TequilaLLM 1B 1.67 0.645 0.305 0.391 0.710 0.542 0.519 SherryLLM 1B 1.25 0.647 0.309 0.388 0.699 0.550 0.519 LLaMA3.2 3B 16 0.745 0.422 0.552 0.768 0.691 0.636 TernaryLLM ∗ 3B 1.67 0.361 0.161 0.260 0.572 0.496 0.370 ParetoQ ∗ 3B 1.67 0.498 0.231 0.303 0.645 0.529 0.441 LLM-QAT 3B 1.67 0.445 0.307 0.434 0.627 0.506 0.464 BitNet 3B 1.67 0.614 0.283 0.429 0.715 0.593 0.527 Spectra 3.9B 1.67 0.660 0.319 0.483 0.744 0.631 0.567 TequilaLLM 3B 1.67 0.702 0.346 0.464 0.739 0.627 0.576 SherryLLM 3B 1.25 0.688 0.364 0.452 0.736 0.593 0.567 Table 2: Comparison of SherryLLM with LLaMA-based ternary LLMs across different model sizes; ∗ indicates results obtained from our reproduction. Figure 6: Ablation study of Arenas. 4.2 Performance Evaluation Comparison of Ternary Quantization Methods: As shown in Table 1 , Sherry achieves performance parity with the current SOTA, Tequila, across both 1B and 3B scales, despite utilizing a significantly lower bit-width of 1.25 bits. On the 1B model, Sherry matches the average SOTA accuracy exactly, effectively matching the 1.67-bit baseline while achieving a 25% reduction in bit-width. Notably, on reasoning-intensive benchmarks like ARC-Challenge, Sherry even outperforms Tequila and narrows the gap to the full-precision BF16 baseline to less than 0.5%. These results demonstrate that 3:4 structured sparsity preserves high-level linguistic capabilities, while the Arenas module successfully addresses optimization trapping. This synergy enables a high-efficiency packing strategy that simultaneously ensures superior hardware alignment and competitive model quality. Comparison of Ternary LLMs: To further evaluate the effectiveness of Sherry, we denote the resulting models as SherryLLM and compare them against existing ternary LLaMA-based architectures. For a fair comparison, we reproduced methods with available training code Liu et al. 2025 ; Chen et al. 2024b on 10B tokens from the UltraFineWeb dataset using identical hyperparameters, while reporting original publication results for others. As shown in Table 2 , SherryLLM achieves at least second-best average accuracy across both scales despite its significantly lower bit-width of 1.25 bits. These results confirm that the integration of 3:4 structured sparsity with the Arenas module preserves model expressivity, allowing SherryLLM to achieve a 25\% bit-width reduction while maintaining competitive performance against baselines. Granularity Average Acc \pm Std Per-tensor 0.502 \pm 0.010 Per-channel 0.513 \pm 0.011 Per-group 0.519 \pm 0.014 Table 3: Average accuracy of Sherry across various quantization granularities. Figure 7: The schedules of annealing factor \lambda_{t} . Figure 8: Comparison of different schedules for \lambda_{t} . Impact of Quantization Granularity: Quantization granularity involves a critical trade-off between hardware efficiency and representational accuracy. While coarse-grained per-tensor quantization maximizes acceleration, it often introduces significant error, whereas fine-grained per-group strategies mitigate this at the cost of increased memory overhead for scaling factors. We evaluate Sherry across per-tensor, per-channel, and per-group (size 128) granularities on five benchmarks. As the results of average accuracy shown in Table 3 , Sherry maintains robust performance across all levels with minimal degradation. This stability is primarily driven by the Arenas module, which provides a continuous gradient flow that allows latent weights to adapt to varying scaling constraints. Effectiveness of Arenas: To formally demonstrate the effectiveness of our training framework, we conducted an ablation study across three quantization schemes: binary (1-bit), 3:4 structured sparse (1.25-bit), and pure ternary AbsMean (1.67-bit). As shown in the Fig 6 , the integration of Arenas yields consistent performance improvements across all configurations. Notably, all the schemes exhibit significant gains. We attribute this to the Arenas’s ability to mitigate the trapping issue in both 1-bit and 1.67-bit, as visualized in Fig. 10 . Inference Efficiency: To empirically validate the efficiency of Sherry, we evaluate the token generation speed against 1.67-bit (TL2) and 2-bit (I2_S) baselines within the BitNet.cpp framework ( Wang et al. 2025b ) using per-channel quantization, alongside a BF16 baseline. Experiments were conducted on an Intel i7-14700HX CPU using 700M and 3B BitNet variants. Both Sherry and the baselines utilize the BitNet.cpp paradigm. The results in Table 4 demonstrate that Sherry outperforms 1.67-bit and 2-bit variants. Specifically, for the 3B model, Sherry achieves a 18\% speedup over the 1.67-bit baseline. This improvement is attributed to our 3:4 structured sparsity and hardware-aligned 5-bit packing, which maximizes SIMD vector lane utilization and eliminates the bit-shuffling overhead inherent in non-power-of-two packing schemes. Scale Method Bits Speed (t/s) \uparrow Size(MB) \downarrow 0.7B BF16 16 34.01 1360.0 I2_S 2.0 132.13 256.56 TL2 1.67 116.83 233.44 Sherry 1.25 148.27 205.50 3B BF16 16 7.55 6190.0 I2_S 2.0 41.87 873.65 TL2 1.67 38.80 846.01 Sherry 1.25 45.55 712.40 Table 4: Inference efficiency on Intel i7-14700HX. Impact of \lambda_{t} Schedules: To evaluate the sensitivity of the annealing process, we compare three decay schedules for the annealing gate \lambda_{t} : linear, exponential, and cosine, alongside their warmup counterparts. As shown in Fig. 8 , every schedule consistently outperforms the baseline without Arenas, demonstrating the robust effectiveness of the Arenas regardless of the specific decay curve. Notably, warmup can increase performance for all schedules. We attribute it from early-stage optimization dynamics in warmup: by gradually introducing the residual, the model establishes a stable foundation before the annealing influence peaks. 5 Conclusion In this work, we introduce Sherry, a novel ternary quantization framework designed to resolve the trade-off between bit width and inference speed. By leveraging a 3:4 structured sparsity pattern, Sherry achieves a 1.25-bit-width that natively aligns with SIMD vector lanes, effectively resolving the hardware under-utilization issues common in previous ternary packing strategies. To combat the performance degradation due to weight trapping, we proposed the Arenas module. Arenas provides an annealing, heterogeneous gradient flow during the training phase, allowing sparse models to retain the expressive diversity. Our extensive evaluations on the LLaMA-3.2 models demonstrate that Sherry achieves performance parity with SOTA ternary models while utilizing 25% fewer bits and 10\% speed up. Limitation While Sherry advances the efficiency of edge-deployed LLMs, several limitations offer avenues for future research: Edge-Centric Model Scale: Our evaluation focuses on models up to 3B parameters, as these are the primary candidates for local edge deployment. While we demonstrate that Sherry achieves a superior Pareto frontier for these scales, the behavior of the Arenas mechanism and the 3:4 sparsity pattern on larger, server-grade models (70B+) remains to be validated. Lack of Server-Specific Optimizations: To prioritize edge inference efficiency, our experiments focus on general SIMD-aligned vector packing rather than server-specific hardware optimizations, such as NVIDIA Sparse Tensor Cores. While this does not affect Sherry’s utility for mobile and local processors, future benchmarks on data-center GPUs could further explore its potential for high-throughput server applications. Weight-Only Quantization: To maximize weight-streaming throughput on edge devices, this study focuses on weight-only ternary quantization. While 1.25-bit weights drastically reduce the static memory footprint, the activations and KV-cache remain in BF16. Future integration with activation quantization could further alleviate memory bottlenecks during long-context inference. Training Overhead: Sherry is designed for maximum inference-phase efficiency. However, the Arenas mechanism increases the compute overhead during the training (QAT) phase. While this is a one-time cost that does not affect edge deployment, it is a factor to consider for researchers with restricted training budgets. Figure 9: The LUT-Based Inference engine for Sherry . The inference engine constructs a dynamic lookup table from input activations; weight indices are then used to fetch pre-computed values. Appendix Appendix A Inference Engine Design Sherry seamlessly integrates with the BitNet.cpp Wang et al. 2025b framework to enable efficient, multiplication-free inference. As illustrated in Fig. 9 , our engine operates in two distinct phases: an offline packing stage and an efficient online inference stage. Offline Packing Phase During the offline phase, the 3:4 sparse ternary weights are compressed into a hardware-aligned format consisting of index and sign components. To maximize SIMD efficiency and memory density, every four-element block is packed into a 5-bit metadata structure: a 4-bit index represents the magnitude/sparsity pattern within the block, and a 1-bit value represents the shared or dominant sign. This packing strategy ensures that the weights are aligned with standard word boundaries, significantly reducing the bit-shuffling overhead compared to non-power-of-two schemes. Online Inference Phase At inference time, the engine utilizes a high-performance lookup table (LUT) paradigm. The input activations are preprocessed into local LUTs for each weight segment. For each segment, the corresponding weight index serves as a direct pointer to retrieve pre-computed results from the LUT, entirely replacing floating-point multiplications with memory-efficient lookups. Following retrieval, the sign weight is applied to determine the final polarity of the segment sum. Subsequent accumulation across segments is performed via integer addition, concluding with the addition of the channel-wise scaling factor \alpha . This architecture results in a highly optimized inference path that realizes the theoretical efficiency of 1.25-bit quantization while requiring only minimal modifications to existing low-bit inference kernels. Appendix B Related Work B.1 Quantization for LLMs Quantization has emerged as a cornerstone technique for enhancing the efficiency of Large Language Models (LLMs) by reducing the bit-precision of weights and activations ( Dettmers et al. 2021 ; Dettmers et al. 2022 ; Lin et al. 2023b ; Frantar et al. 2022 ; Lin et al. 2025 ; Yang et al. 2025 ; Li et al. 2023 ) . While effective, popular weight-only quantization methods ( Lin et al. 2023b ; Frantar et al. 2022 ) typically necessitate mixed-precision matrix multiplication, where weights and activations reside in disparate data formats. This heterogeneity requires specialized hardware support to maintain throughput, posing a significant hurdle for deployment on diverse edge and mobile platforms where hardware specialized for non-standard bit-widths is often unavailable. To address this, weight-activation quantization strategies ( Dettmers et al. 2022 ; Xiao et al. 2023 ; Huang and Wu 2025 ; Wei et al. 2022 ) seek a unified low-precision format. However, these methods often encounter the "outlier problem" ( Xiao et al. 2023 ; Dettmers et al. 2022 ) , where extreme activation values lead to high quantization errors. Consequently, it remains challenging for activations to reach the same ultra-low precision as weights without severe performance degradation, leaving a gap in achieving truly hardware-agnostic, high-efficiency deployment on the edge. B.2 Ternary Quantization Ternary quantization Li et al. 2016 ; Zhu et al. 2016 , often referred to as 1.58-bit quantization, offers a paradigm shift by constraining weights to the set \{-1,0,+1\} . Beyond substantial memory reduction, this approach fundamentally simplifies the core matrix multiplication operation into addition, effectively eliminating the need for power-intensive multipliers Ma et al. 2025 ; Wang et al. 2023 . This intrinsic hardware-friendliness is particularly vital for resource-constrained mobile environments. Early research in this domain primarily focused on optimizing the quantization function through thresholding and scaling. Ternary Weight Networks (TWN) ( Li et al. 2016 ) minimized reconstruction distortion by assuming a Gaussian weight distribution, while Trained Ternary Quantization (TTQ) ( Zhu et al. 2016 ) introduced learnable scaling factors to adapt the ternary representation during training. Subsequent work by Leng et al. 2018 leveraged the Alternating Direction Method of Multipliers (ADMM) to iteratively optimize these parameters. The advent of Large Language Models (LLMs) Wu et al. 2023 ; Floridi and Chiriatti 2020 ; Zhang et al. 2022 has revitalized interest in ternary schemes, as the generative capacity of LLMs is notoriously sensitive to precision loss. This has led to two distinct research trajectories. The first utilizes Post-Training Quantization (PTQ) Lin et al. 2023b ; Frantar et al. 2022 ; Xiao et al. 2025 to minimize the computational cost of quantization; however, PTQ often suffers from non-negligible performance drops at ultra-low bit-widths. Consequently, Quantization-Aware Training (QAT) Chen et al. 2025 has become the preferred choice for robust performance recovery. Within the QAT landscape, strategies range from the straightforward AbsMean scaling utilized by the BitNet family Ma et al. 2025 ; Wang et al. 2023 ; Wang et al. 2025b to more sophisticated adaptations of Learned Step Size Quantization (LSQ) Esser et al. 2019 for the ternary regime Chen et al. 2024b ; Liu et al. 2025 . Regarding the inference engine, classical frameworks like llama.cpp still rely on multiplication-based kernels that require weight unpacking and floating-point operations. To overcome this, T-Mac Wei et al. 2025 and TENET Huang et al. 2025c proposed a lookup table (LUT)-based engine to eliminate multiplications, though it requires packing ternary weights into 2-bit containers. BitNet.cpp Wang et al. 2025b attempted to reduce the footprint further via a 1.67-bit packing strategy (3 weights in 5 bits). However, this 3-way packing is inherently SIMD-unfriendly, introducing significant bit-shuffling overhead that often results in slower performance than 2-bit packing. Consequently, existing ternary models are forced to trade off between bit width and inference speed. Sherry addresses this dilemma by introducing hardware-aligned 3:4 structured sparsity, achieving a 1.25-bit SIMD-friendly packing. Figure 10: Weight distributions across different quantization regimes. The inclusion of Arenas effectively mitigates the trapping phenomenon in both binary and naive ternary quantization. B.3 N:M sparsity N:M structured sparsity Lin et al. 2023a ; Zhou et al. 2021 has emerged as a critical middle ground between the flexibility of unstructured pruning and the hardware efficiency of block-wise pruning. By enforcing that exactly N out of every M contiguous weights are non-zero, this pattern provides predictable memory access patterns highly amenable to hardware acceleration. Notably, NVIDIA’s Ampere and subsequent architectures Lin et al. 2023a introduced native Tensor Core support for 2:4 sparsity, doubling throughput with minimal accuracy loss in high-precision models. Existing research Sun et al. 2021 ; Fu et al. 2023 ; Zhang et al. 2023 on N:M sparsity has primarily focused on the selection of optimal masks and performance recovery via specialized training regimes. However, most efforts are not coordinated with ultra-low bit quantization, as they are largely designed for Sparse Tensor Cores on GPUs, which currently prioritize 16-bit or 32-bit floating-point arithmetic. In the context of ultra-low bit-widths, the intersection of N:M sparsity and ternary quantization remains largely unexplored. Furthermore, as the bit-width decreases, the rigid N:M constraint exacerbates the weight trapping phenomenon. Sherry extends the N:M paradigm by introducing a hardware-aligned 3:4 pattern coupled with the Arenas module. This synergy is specifically designed to bridge the accuracy gap between sparse and dense ternary models with zero additional inference overhead. To our knowledge, Sherry is the first to provide a hardware-efficient solution for applying N:M sparsity on ultra-low bit-width quantization. Appendix C Optimality of the 3:4 Sparse Format We argue that the 3:4 structured sparsity pattern represents the optimal solution among all N:M ( N0 . Index Optimality: Let \{|W|_{(1)},|W|_{(2)},|W|_{(3)}, |W|_{(4)}\} be the sorted absolute values of weights in a 4-element block such that |W|_{(1)}\geq|W|_{(2)}\geq|W|_{(3)}\geq|W|_{(4)} . The correlation sum for the block is \sum_{k\in\mathcal{I}}|W|_{(k)} , where \mathcal{I} is the set of indices of the 3 chosen elements. To maximize this sum, we must choose the indices corresponding to the three largest magnitudes: (1),(2), and (3) . By combining the sign and index selection, for each block i\in[b,b+3] , the optimal ternary element T_{i,j} is given by: T^{*}_{i,j}=\begin{cases}0,&\text{if }i=\arg\min|W_{i,j}|;\\ \text{sign}(W_{i,j}),&\text{otherwise}.\end{cases} (12) Substituting T_{i,j}^{*} back into the expression for \alpha_{j}^{*} gives the global minimum: \alpha^{*}_{j}=\frac{4}{3d_{in}}\sum_{i\in\mathcal{S}_{j}}|W_{i,j}|, (13) where \mathcal{S}_{j}=\{i\mid T_{i,j}\neq 0\} denotes the set of active (non-zero) indices in the j -th column. Appendix E Ternary Quantization Baselines Recall the general form of ternary quantization: taking per-channel quantization as an example, for a full-precision weight matrix W\in\mathbb{R}^{d_{in}\times d_{out}} , the general ternary quantization function Q(\cdot) is defined as: Q(W)=T\alpha,\quad T_{i,j}=\begin{cases}+1,&\text{if }W_{i,j}>\Delta_{j};\\ 0,&\text{if }|W_{i,j}|\leq\Delta_{j};\\ -1,&\text{if }W_{i,j}<-\Delta_{j},\end{cases} (14) where T\in\{-1,0,+1\}^{d_{in}\times d_{out}} is the ternary weight matrix, \alpha\in\mathbb{R}^{d_{out}} represents the scaling factors, and \Delta\in\mathbb{R}^{d_{out}} denotes the quantization thresholds. Figure 11: Weight distributions and the Effective Ranks (ER) of their gradients across different layers. For contemporary large-scale ternary models ( Ma et al. 2025 ; Team et al. 2025 ) , the AbsMean method has emerged as the de facto standard due to its superior training stability. In AbsMean, the scaling factor \alpha_{j} and threshold \Delta_{j} for the j -th channel are calculated as: \alpha_{j}=\frac{1}{d_{in}}\sum_{i=1}^{d_{in}}|W_{i,j}|,\quad\Delta_{j}=\frac{\alpha_{j}}{2}. (15) Besides Absmean, previous optimization strategies can be broadly categorized into: (1) reducing quantization error and (2) enhancing model expressive capacity. E.1 Reducing Quantization Error A primary line of work focuses on optimizing \Delta and \alpha to minimize the reconstruction error \min_{\Delta,\alpha}\|W-T\alpha\|^{2}_{2}. (16) This is exemplified by estimation-based methods like Ternary Weight Networks (TWN) ( Li et al. 2016 ) , which minimize distortion by assuming a Gaussian weight distribution to approximate \Delta^{*}_{j}\approx 0.7\cdot\mathbb{E}[|W_{:,j}|]. (17) For a fixed \Delta , the optimal \alpha is: \alpha^{*}_{j}=\frac{1}{|\mathcal{S}_{j}|}\sum_{i\in\mathcal{S}_{j}}|W_{i,j}|, (18) where \mathcal{S}_{j} is the set of weights W_{i,j} exceeding \Delta_{j} . However, the Gaussian assumption often fails in LLMs, leading to biased estimates and performance degradation. While subsequent methods like LSQ ( Esser et al. 2019 ) and DLT ( Chen et al. 2024b ) introduce trainable scaling factors, they are easier to being trapped in sub-optima, exhibiting slower convergence and higher final loss compared to AbsMean Huang et al. 2025a . E.2 Enhancing Expressive Capacity Alternative approaches attempt to recover capacity by incorporating bias terms. DLT Chen et al. 2024b introduces a learnable bias during dequantization: Y=X(T\alpha+B)=XT\alpha+Xb. (19) However, this necessitates a dense full-precision multiplication ( Xb ), destroying the computational efficiency of ternary quantization. Similarly, SEQ Liu et al. 2025 reassigns the zero-point to a non-zero value \alpha_{j}b_{j} : T_{i,j}=\begin{cases}+1,&\text{if }W_{i,j}>\Delta_{j};\\ \alpha_{j}b_{j},&\text{if }|W_{i,j}|\leq\Delta_{j};\\ -1,&\text{if }W_{i,j}<-\Delta_{j},\end{cases} (20) This also voids hardware efficiency, as the resulting operations are no longer multiplication-free. In contrast, Sherry maintains strict ternary efficiency during inference by utilizing an annealing residual that vanishes post-training. Appendix F Effective Rank for Gradient Analysis To quantitatively assess the diversity and information density of the gradient updates, we utilize the Effective Rank ( ER ) metric Roy and Vetterli 2007 . While the standard algebraic rank provides a binary measure of linear independence, it is highly sensitive to infinitesimal noise and fails to capture the numerical stability of the gradient matrix. In contrast, Effective Rank provides a continuous measure of the "learning dimensionality" of a layer by evaluating the entropy of its singular value distribution. Given a weight gradient matrix G\in\mathbb{R}^{d_{in}\times d_{out}} , let \sigma=(\sigma_{1},\sigma_{2},\dots,\sigma_{k}) denote its singular values obtained via Singular Value Decomposition (SVD), sorted in descending order. We first normalize these values to form a probability-like distribution p : p_{i}=\frac{\sigma_{i}}{\sum_{j=1}^{k}\sigma_{j}},\quad i=1,\dots,k. (21) The Effective Rank is then defined as the exponential of the Shannon entropy of this distribution: ER(G)=\exp\left(-\sum_{i=1}^{k}p_{i}\ln p_{i}\right). (22) The value of ER(G) ranges from 1 to k . An ER close to 1 indicates that the gradients are highly homogenized , with the update signal collapsing into a single dominant direction. Conversely, an ER close to k suggests a high-rank, diverse update where many independent features are being learned simultaneously. In the context of N:M sparsity, we use ER to diagnose Gradient Homogenization . As shown in Figure 5 and 11 , naive 3:4 training regimes often exhibit a relatively small ER as weights become "trapped" in discrete slots, leading to redundant updates. By maintaining a higher ER , Sherry ensures that the sparse ternary model retains the representational richness of its dense counterpart. Appendix G More Experimental Details and Results G.1 Evaluation Benchmarks PIQA: The Physical Interaction Question Answering (PIQA) benchmark Bisk et al. 2020 focuses on physical commonsense reasoning. It tests a model’s understanding of everyday physical laws by presenting scenarios (e.g., "How do you stabilize a wobbly table?") and requiring the selection of the correct solution from two options. Success on PIQA indicates a foundational grasp of physical object interactions. ARC-Easy and ARC-Challenge: The AI2 Reasoning Challenge (ARC) dataset Clark et al. 2018 is bifurcated to assess scientific knowledge. The ARC-Easy set contains grade-school science questions often answerable via fact retrieval. Conversely, ARC-Challenge consists of questions curated to be difficult for standard retrieval algorithms, requiring complex multi-step reasoning and a deeper conceptual understanding. HellaSwag: HellaSwag Zellers et al. 2019 evaluates contextual commonsense inference. Models are given a situational premise and must select the most plausible continuation from four options. Crucially, distractors are adversarially generated to be deceptively plausible, ensuring that success requires a nuanced understanding of event dynamics rather than simple word association. WinoGrande: WinoGrande Sakaguchi et al. 2021 is a large-scale dataset for assessing commonsense reasoning through pronoun resolution. Derived from the Winograd Schema Challenge, it presents sentences with ambiguous pronouns and requires the model to determine the correct referent. WinoGrande employs adversarial filtering to reduce statistical biases, forcing models to rely on genuine commonsense understanding. G.2 Annealing Gate Schedule \lambda_{t} The hyperparameter \lambda_{t} plays a crucial role in the annealing gate. Proper scheduling of \lambda_{t} from an initial value of 1 to a final value of 0 throughout the training process can significantly improve model performance and convergence stability. This section describes three decay strategies in Fig. 8 for \lambda_{t} scheduling, where p_{t}\in[0,1] represents the training progress (0: training start, 1: training completion). The linear decay strategy provides a constant reduction rate of \lambda_{t} throughout the training process: \lambda_{t}=1-p_{t}. (23) The cosine decay strategy follows a convex curve, with slower decay at the beginning and end of training, and faster decay in the middle stage: \lambda_{t}=\frac{1}{2}\left[1+\cos(\pi p_{t})\right]. (24) The exponential decay strategy follows a concave curve, with rapid decay at the beginning that gradually slows down: \lambda_{t}=\exp(-5p_{t}) (25) Appendix H Broader Impact The development of Sherry carries significant implications for the democratization and sustainability of Large Language Models (LLMs). By achieving high-performance 1.25-bit inference through hardware-aligned 3:4 structured sparsity, our work opens up significant societal and technological benefits in resource-constrained scenarios Li et al. 2025 ; Huang et al. 2023 ; Huang et al. 2024a , such as edge/mobile computing Forman and Zahorjan 1994 ; Imielinski and Korth 1996 ; Chen et al. 2024a and cross-device federated learning McMahan et al. 2017 ; Huang et al. 2025b . This empowers researchers and developers in resource-constrained environments and developing regions to access and innovate with large-scale models without prohibitive hardware costs.