AI News AI资讯 1d ago Updated 22h ago 更新于 22小时前 50

PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B That Run on Laptops and Phones PrismML发布Bonsai 27B:可在笔记本和手机上运行的Qwen3.6-27B的1位和三元构建版本

PrismML releases Bonsai 27B, a quantized version of Qwen3.6-27B using ternary (1.71 bits/weight) and binary (1.125 bits/weight) representations without retraining. The ternary variant retains 94.6% of the FP16 baseline performance in a compact 5.9GB footprint, while the 1-bit variant retains 89.5% at 3.9GB. Unlike conventional sub-4-bit quantizations that suffer selective collapse on complex reasoning tasks, Bonsai maintains robust performance on math, coding, and agentic benchmarks. The archite PrismML发布Bonsai 27B,将Qwen3.6-27B量化为1.71位三元组和1.125位二进制版本,无需重新预训练。 三元组版本保留94.6%的FP16性能,体积仅5.9GB;1位版本保留89.5%性能,体积仅3.9GB。 通过线性注意力机制和4位KV缓存优化,有效解决了长上下文(262K tokens)下的内存瓶颈。 相比传统亚4位量化,Bonsai在复杂推理、代码和代理任务上避免了性能崩塌,更适合端侧部署。 配合DSpark推测解码技术,在H100上可实现1.37倍加速,且验证过程无损,输出分布一致。

68
Hot 热度
72
Quality 质量
75
Impact 影响力

Analysis 深度分析

TL;DR

  • PrismML releases Bonsai 27B, a quantized version of Qwen3.6-27B using ternary (1.71 bits/weight) and binary (1.125 bits/weight) representations without retraining.
  • The ternary variant retains 94.6% of the FP16 baseline performance in a compact 5.9GB footprint, while the 1-bit variant retains 89.5% at 3.9GB.
  • Unlike conventional sub-4-bit quantizations that suffer selective collapse on complex reasoning tasks, Bonsai maintains robust performance on math, coding, and agentic benchmarks.
  • The architecture leverages linear attention for ~75% of operations and a separate 4-bit vision tower, enabling a 262K token context window feasible for mobile and edge deployment.
  • Integration with DSpark speculative decoding provides a 1.37x throughput speedup on H100s, making the models highly efficient for both local inference and cloud serving.

Why It Matters

This release demonstrates that extreme low-bit quantization (1-2 bits) can preserve the capabilities of large language models significantly better than traditional methods, challenging the assumption that such compression leads to inevitable performance collapse. For AI practitioners, it offers a viable path to deploying 27B-class multimodal models on consumer hardware like iPhones and laptops, bridging the gap between high-performance cloud models and resource-constrained edge devices.

Technical Details

  • Quantization Method: Uses a group-wise scaling approach where each weight is represented as $w_i = s_g \cdot t_i$, with one FP16 scale per 128 weights. Ternary weights use ${-1, 0, +1}$ (1.71 bpw), and binary weights use ${-1, +1}$ (1.125 bpw).
  • Architecture Preservation: The base Qwen3.6-27B architecture remains unchanged, including a ~24.8B language model, 0.46B vision tower (held at 4-bit HQQ), and 2.5B embeddings/head. Approximately 75% of attention is linear to manage KV cache growth.
  • Performance Benchmarks: Evaluated on 15 benchmarks including AIME26, LiveCodeBench, and MMLU-Redux. Ternary Bonsai scores 80.49 average thinking score vs 85.07 for FP16, significantly outperforming IQ2_XXS (72.73) which collapses on complex reasoning tasks.
  • Deployment Efficiency: The 1-bit model fits within the ~6GB usable memory of a 12GB iPhone due to iOS app memory limits. On H100 GPUs, it achieves 104.8 tokens/sec in generation and 2755 tokens/sec in prefill, with speculative decoding boosting generation to 143.8 tokens/sec.

Industry Insight

  • Edge AI Viability: The ability to run 27B-class models on mobile devices with minimal battery impact (672 tokens per 1% battery) signals a shift toward on-device AI agents, reducing reliance on cloud infrastructure for latency-sensitive or privacy-critical applications.
  • Quantization Strategy Shift: Developers should reconsider standard sub-4-bit quantizations for complex reasoning tasks, as they may hide catastrophic failures in agentic or mathematical benchmarks. Specialized low-bit formats like Bonsai offer a safer alternative for maintaining capability density.
  • Hardware-Aware Optimization: The success of linear attention and specific KV cache management strategies highlights the importance of co-designing model architectures with deployment constraints, particularly for mobile and edge scenarios where memory bandwidth and capacity are strict binding constraints.

TL;DR

  • PrismML发布Bonsai 27B,将Qwen3.6-27B量化为1.71位三元组和1.125位二进制版本,无需重新预训练。
  • 三元组版本保留94.6%的FP16性能,体积仅5.9GB;1位版本保留89.5%性能,体积仅3.9GB。
  • 通过线性注意力机制和4位KV缓存优化,有效解决了长上下文(262K tokens)下的内存瓶颈。
  • 相比传统亚4位量化,Bonsai在复杂推理、代码和代理任务上避免了性能崩塌,更适合端侧部署。
  • 配合DSpark推测解码技术,在H100上可实现1.37倍加速,且验证过程无损,输出分布一致。

为什么值得看

这篇文章展示了极低比特量化(1-2 bit)在保持大型多模态模型性能方面的最新突破,证明了无需从头预训练即可实现高效压缩。对于希望将27B级别高性能模型部署到手机、笔记本等边缘设备的开发者和企业而言,Bonsai提供了切实可行的技术路径和性能基准。

技术解析

  • 量化方案与架构:Bonsai 27B基于Qwen3.6-27B架构,采用分组缩放策略(每128个权重共享一个FP16缩放因子)。三元组版本使用{-1, 0, +1}权重(1.71 bits/weight),1位版本使用{-1, +1}权重(1.125 bits/weight)。视觉塔单独保持4-bit精度(HQQ)。
  • 内存与上下文优化:模型利用约75%的线性注意力机制来降低KV缓存开销。结合4位KV缓存,262K token上下文所需的内存从FP16的17.2GB降至4.3GB,使得在单张24GB GPU上运行成为可能。
  • 性能对比:在15项基准测试中,三元组版本平均得分80.49(基线85.07),1位版本76.11。传统2-bit量化(如IQ2_XXS)在AIME26等复杂任务上得分骤降至57.5,而Bonsai保持了较高的鲁棒性。
  • 推理加速:引入了针对Bonsai训练的DSpark草稿模型。在H100上,1位版本在推测深度k=4时达到平均接受长度τ=3.6,吞吐量提升至143.8 tok/s(加速1.37倍),且验证阶段无损。

行业启示

  • 端侧大模型部署新标准:1位/1.7位量化使得27B参数量的多模态模型能够装入智能手机(如iPhone 17 Pro Max),标志着“本地优先”的AI代理应用进入实用阶段,隐私保护和离线工作能力得到显著提升。
  • 量化技术的演进方向:简单的后训练量化(PTQ)在极低比特下容易导致特定能力(如逻辑推理、工具调用)崩塌。Bonsai的成功表明,针对特定架构(如线性注意力)优化的量化方法和混合精度策略是未来降低显存占用并保持性能的关键。
  • 硬件效率与软件栈协同:通过减少内存带宽压力(生成阶段受限于内存带宽),极低比特模型能显著提升吞吐量。这提示开发者在部署低比特模型时,应重点关注内存带宽利用率及推测解码等推理优化技术,以最大化硬件性能。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Multimodal 多模态 Quantization 量化 Deployment 部署