Qwen releases FP8 27B vision-language model with 262K native context and agent benchmark gains
- Qwen3.8-27B-FP8 is a 27B-parameter dense vision-language model released in Hugging Face Transformers format, using fine-grained FP8 quantization with 128-token blocks that Qwen says has near-identical performance to the original model.
- Qwen reports 61.7 on SWE-bench Pro, 79.0 on QwenSWEBench, and 42.2 on DeepSWE 1.1, versus 53.5, 49.3, and 13.3 respectively for Qwen3.6-27B under its stated evaluation setups.
- The model has 262,144 native context tokens and is extensible to 1,000,000 tokens; the upcoming hosted Qwen Cloud version is planned to use a 1M-token context by default.
- Thinking mode is enabled by default but can be disabled per request, while
reasoning_effortcontrols reasoning depth andpreserve_thinkingretains reasoning context across prior messages. - Qwen reports native image and video understanding, including STEM diagrams, documents, and hour-scale video, and lists compatibility with Transformers, vLLM, SGLang, and TokenSpeed.
Hacker News opinions
I saw Qwen 3.6 27B as the best size-to-capability compromise for consumer hardware, so I am eager to see what people build with this release.
I would not equate benchmark wins with real-world parity with Opus. Qwen models are good enough for many local tasks, but people who use them like API frontier models can end up disappointed.
A 27B model scoring 42.2 on DeepSWE versus 40 for Opus 4.7 Max is hard to believe given the size gap. I want to run my own tests before accepting that result.
I wish every GGUF quantization had the original model's benchmark suite rerun. Without that, it is hard to compare the accuracy cost of each quant.
I get about 30 to 35 tokens per second on an RTX 3090 without MTP and 60 to 70 with it. Without multi-token prediction, I would rather pay for a fast API than use my local setup.
My Q4_K_M run gets about 48 tokens per second on an RTX 4090 with --flash-attn on --parallel 1 --load-mode mmap. Speculative decoding should push dual 3090s past 100 tokens per second.
I want a 35B A3B-style MoE version. Qwen 3.5 122B A10B runs about twice as fast as this 27B dense model, and Qwen's 35B A3B model is much more practical on an iGPU or MacBook.
I am sticking with Gemma 4 26B A3B for now. Qwen 3.8 seems prone to overthinking, while Gemma runs about four times faster and often uses far fewer reasoning tokens for a similar answer.
I would wait for llama.cpp to catch up before judging it. In this case, though, the architecture appears unchanged from Qwen 3.5 and 3.6, so llama.cpp should already understand it.
I wanted Gemma 4 26B in the comparison table. Comparing dense 27B with a 26B MoE is imperfect, but it would still help people choosing a local model.
The benchmark numbers are huge, but I am worried about benchmark optimization. If Qwen claims Opus 4.6-level performance, it should at least hold up against Sonnet 4.6 in ordinary use.