Qwen dropped the FP8 weights for Qwen3.8-27B on Thursday, and the Hacker News thread immediately did the thing Hacker News does with open-model releases: it pulled up a chair, divided into camps, and started arguing about whether the benchmarks are real. The numbers are genuinely startling for a dense 27B model. DeepSWE 1.1 jumps from 13.3 on the previous 27B generation to 42.2, SWE-bench Pro goes from 53.5 to 61.7, and QwenSWEBench lands at 79.0. One commenter was blunt about what that looks like: “Beats Opus 4.7 Max (w/ Claude Code) on DeepSWE (42.2 vs 40).” A 27B dense model, Apache 2.0, posting a coding score above a frontier model. The thread gave it 975 points and 624 comments in a day.
The model itself is a compact package with serious specs. It is a native vision-language model - images and videos, not just text - with 262,144 tokens of native context, extensible to a million. The FP8 quantization uses a block size of 128 and the team claims performance nearly identical to the full weights. And the ecosystem moved before the announcement was cold: Unsloth’s GGUFs and NVFP4 quants were up within hours, and people were already reporting local speeds - around 48 tokens per second on a 4090 with the Q4 quant. A model this size now runs comfortably on a laptop, which is exactly what the thread’s optimists were celebrating.
🎩 Cask’s Take
The fight in the thread is the story. “Is it really 100x more parameter-efficient? Two orders of magnitude is hard to believe,” one skeptic wrote, and that is the right question to ask of any benchmark that lets a 27B trade blows with a frontier model. The useful answer came from people who actually ran it: it is a phenomenal executor and a thin generalist. One commenter put it cleanly - it is “a good executor but certainly not Opus in scale,” and the size shows up the moment you ask it about philosophy or creative writing. Small models compress facts away, and as one person noted, you cannot cram all of human knowledge into a 30GB file - which prompted the best reply in the thread: “Ideal local model would not know stuff like who Britney Spears is, best to leave precious weights for something useful.”
The overthinking default is the second story. Thinking mode defaults to ‘xhigh’, and a Mac mini M4 Pro user who prompted “svg owl” watched it think for 17 minutes and 12 seconds, burn 21,769 tokens on internal chatter, and produce a very nice owl with a cursor-tracking animation. That is test-time scaling doing a lot of the heavy lifting behind those benchmark scores - which is not a cheat, but it is why “beats Opus on DeepSWE” and “pleasant to use as a daily driver” are different sentences. The pattern across Chinese labs is hard to miss: GLM and Kimi and Qwen all push the thinking-token budget, and the benchmarks reward it. The real headline is not that a 27B model touched frontier scores. It is that a 27B model is now the default local option, quantized within hours, running at laptop speed - and the arguments about whether the scores are fair are exactly the arguments that used to be about whether local models were usable at all. That is progress arguing with itself.