Qwen3.8 27B
An open-source lightweight multimodal large language model released by Alibaba's Qwen team, with 27 billion parameters. It supports native multimodal capabilities including text and image understanding, a native context window of 260,000 tokens, and is released under the Apache 2.0 license.
Core Facts
Timeline (last 90 days)
RunInfra does not train models itself; instead it integrates open-source models such as GLM 5.3, DeepSeek V4.1 Flash and Qwen3.8 27B.
The three-tier storage approach is only applicable to MoE models and does not work with Dense models (such as Qwen3.8 27B).
Qwen3.8 27B outperforms GLM 5.2 and some GPT 5.6 models on the agentic index.
Qwen3.8 27B surpasses Opus 4.6 Max in visual performance on computer use and browser operation tasks.
Qwen3.8 27B officially provides a BFloat16 full-precision version and an FP8 version; Unsloth has released a 4-bit version with NVFP4 quantization.
The ablation version of Qwen3.8 27B is prone to falling into an infinite loop of repetitive decoding during the thinking phase, and the blogger does not recommend using it.
The X-high thinking tier of Qwen3.8 27B consumes an extremely large number of tokens (e.g., 11,000 for drawing a pelican and 35,000 for drawing a dragon), while the medium tier is the best cost-performance sweet spot.
The choice of inference tier has a greater impact on model usability than the choice of quantization method.
On an RTX Pro 6000 (96 GB VRAM), the official FP8 version combined with built-in MTP speculative decoding can reach approximately 120 tokens/second, with quality nearly indistinguishable from 16-bit.
SGLang using official NVFP4 weights with D-Spark speculative decoding achieved an average of approximately 173 tokens/second on Blackwell GPUs, approaching 200 for short tasks, while loading a 262k context window.
7 more timeline events
All Facts (15)
The reviewer believes that Qwen3.8 27B is stronger than the previous generation Qwen3.7 in almost all aspects of coding capability.
65%UnverifiedRunInfra does not train models itself; instead it integrates open-source models such as GLM 5.3, DeepSeek V4.1 Flash and Qwen3.8 27B.
50%UnverifiedThe three-tier storage approach is only applicable to MoE models and does not work with Dense models (such as Qwen3.8 27B).
50%UnverifiedSGLang using official NVFP4 weights with D-Spark speculative decoding achieved an average of approximately 173 tokens/second on Blackwell GPUs, approaching 200 for short tasks, while loading a 262k context window.
50%UnverifiedQwen3.8 27B surpasses Opus 4.6 Max in visual performance on computer use and browser operation tasks.
50%UnverifiedQwen3.8 27B outperforms GLM 5.2 and some GPT 5.6 models on the agentic index.
50%UnverifiedThe choice of inference tier has a greater impact on model usability than the choice of quantization method.
50%UnverifiedOn an RTX Pro 6000 (96 GB VRAM), the official FP8 version combined with built-in MTP speculative decoding can reach approximately 120 tokens/second, with quality nearly indistinguishable from 16-bit.
50%UnverifiedThe ablation version of Qwen3.8 27B is prone to falling into an infinite loop of repetitive decoding during the thinking phase, and the blogger does not recommend using it.
50%UnverifiedThe X-high thinking tier of Qwen3.8 27B consumes an extremely large number of tokens (e.g., 11,000 for drawing a pelican and 35,000 for drawing a dragon), while the medium tier is the best cost-performance sweet spot.
50%UnverifiedThe official Anifer project initially only supported a single RTX 5090, and only five official models were registered, focusing on Qwen3.8 27B, Qwen3.6 27B, and Qwen3.6 35B MoE models.
50%UnverifiedThere is a smaller approximately 9GB version of the Qwen3.8 27B model on Hugging Face, but its quality will be degraded.
50%UnverifiedQwen3.8 27B still makes notable errors in complex 3D modeling, such as scoring only 2 out of 10 on an F1 car drifting scene and producing common-sense violations in SVG artwork.
50%UnverifiedIn a frontend design test, Qwen3.8 27B's context window was limited to 32K under the current configuration, resulting in incomplete code generation with bugs such as abnormal page reflections and missing SVG icons.
50%UnverifiedQwen3.8 27B officially provides a BFloat16 full-precision version and an FP8 version; Unsloth has released a 4-bit version with NVFP4 quantization.
50%