Qwen3-27B
An open-source model with 27 billion parameters from Alibaba's Qwen3 series, noted for its cost-effectiveness. It supports local deployment on consumer-grade or mid-range professional hardware, with overall performance approaching frontier-level models.
Timeline (last 90 days)
Across six real-world long-chain agentic build tasks, Qwen3-27B scored 35/60 and delivered working applications, while Bonsai 2 collapsed across the board and produced no runnable pages whatsoever.
On simple tasks with thinking disabled and on long-context tests, Bonsai 2 performed nearly on par with full-precision Qwen3-27B, and the 98% claim largely held up.
Bonsai 2 uses a full retraining starting from the Qwen3-27B architecture with ternary constraints as the objective, rather than applying post-training quantization to existing weights.
Bonsai 2 is a ternary-quantized large model released by Prism ML, retrained from scratch on a Qwen3-27B base, with an overall file size of approximately 6 GB.
Each of Qwen3-27B's 16 full-attention layers has 4 KV heads and a head dimension of 256, occupying 64 KB per token at 16-bit precision.
Unsloth updated the GGUF repository for Qwen3-27B to the dynamic 3.0 recipe, expanding calibration from 45 importance matrix blocks to over 1,200, with a single file capable of using up to 14 different GGML quantization types.
On Apple Silicon, the natively recommended approach is MLX 4-bit affine quantization (group size 64), with a 4-bit MLX build occupying approximately 15 GB of unified memory.
The NVFP4 version of Qwen3-27B shows only marginal loss compared to uncompressed FP16 on the MMLU Pro and GPQA Diamond benchmarks.
The NVFP4 version of Qwen3-27B uses sensitivity-aware mixed precision: feed-forward layers run in FP4 (E2M1) with FP8 block scaling, while attention projections, DeltaNet recurrent layers, the vision encoder, and multi-token prediction heads remain in FP8 or BF16.
On 4×RTX 4090 with prefix caching and DFlash2 speculative decoding both enabled, Qwen3-27B's total throughput drops by approximately 37%.
9 more timeline events
All Facts (18)
Bonsai 2 uses a full retraining starting from the Qwen3-27B architecture with ternary constraints as the objective, rather than applying post-training quantization to existing weights.
60%UnverifiedAcross six real-world long-chain agentic build tasks, Qwen3-27B scored 35/60 and delivered working applications, while Bonsai 2 collapsed across the board and produced no runnable pages whatsoever.
50%UnverifiedOn simple tasks with thinking disabled and on long-context tests, Bonsai 2 performed nearly on par with full-precision Qwen3-27B, and the 98% claim largely held up.
50%UnverifiedBonsai 2 is a ternary-quantized large model released by Prism ML, retrained from scratch on a Qwen3-27B base, with an overall file size of approximately 6 GB.
50%UnverifiedEach of Qwen3-27B's 16 full-attention layers has 4 KV heads and a head dimension of 256, occupying 64 KB per token at 16-bit precision.
50%UnverifiedUnsloth updated the GGUF repository for Qwen3-27B to the dynamic 3.0 recipe, expanding calibration from 45 importance matrix blocks to over 1,200, with a single file capable of using up to 14 different GGML quantization types.
50%UnverifiedThe NVFP4 version of Qwen3-27B uses sensitivity-aware mixed precision: feed-forward layers run in FP4 (E2M1) with FP8 block scaling, while attention projections, DeltaNet recurrent layers, the vision encoder, and multi-token prediction heads remain in FP8 or BF16.
50%UnverifiedThe NVFP4 version of Qwen3-27B shows only marginal loss compared to uncompressed FP16 on the MMLU Pro and GPQA Diamond benchmarks.
50%UnverifiedOn Apple Silicon, the natively recommended approach is MLX 4-bit affine quantization (group size 64), with a 4-bit MLX build occupying approximately 15 GB of unified memory.
50%UnverifiedAt approximately 52 concurrent connections, the system is essentially 100% saturated, generating a maximum of approximately 704 tokens per second, but with latency spiking to 64 seconds.
50%Unverified24GB显存是部署的甜点区，跑4Bit量化后仍有余量，性价比最高
50%Unverified单卡RTX 4090跑4Bit量化的通义千问3-27B约每秒生成49个字
50%Unverified通义千问3-27B满血版需要55.6GB显存，而量化版仅需17GB
50%Unverified通义千问3-27B采用Apache 2.0协议，支持免费商用、自定义权重、图片与视频多模态理解，原生上下文长度达256K
50%Unverified24GB显存跑4Bit量化是通义千问3-27B部署的性价比甜点区，日常使用完全够用
50%Unverified稠密架构意味着模型在推理时会激活全部270亿参数进行计算
50%Unverified本地部署通义千问3-27B每百万Token的运行成本约为2.7美元，对标GPT-4.5级别的云端API每百万Token高达约150美元，两者相差约50倍
50%Unverified阿里在8月开源了通义千问3-27B（Qwen3-27B）稠密架构模型，采用Apache 2.0协议
50%