Update: Qwen3.8-27B lands — Apache-2.0, runs on one 24GB GPU
The dense sibling to this week's 2.4T Max is finally out: 262K context, a big SWE-bench jump over 3.6, and a Q4 build that fits a single 24GB card.
Copy markdown
The 27B the local crowd was waiting on
Qwen3.8's flagship arrived earlier this week as a 2.4T-param Max; the 27B dense model — the one that fits real hardware — landed August 14. It's Apache 2.0, weights on Hugging Face, with official BF16 and FP8 checkpoints at launch.
Coding scores jump hard over 3.6
SWE-bench Pro climbs to 61.7 (from 53.5 on 3.6), Terminal-Bench to 73.0 (from 63.4), and DeepSWE more than triples to 42.2. LiveCodeBench hits 90.3 — strong for a 27B, though DeepSeek V4 Flash (82.7 Terminal-Bench) and Opus 5 (79.2 SWE-bench Pro) still lead the frontier rows.
Runs on one 24GB GPU, with a caveat
The Q4_K_M GGUF is about 17 GiB and plausible on a single 24GB card at moderate context; FP8 wants ~48GB, BF16 needs 80GB or multi-GPU. Watch the KV cache — the native 262K context adds roughly 16 GiB on its own.
262K context native, 1M with YaRN
It ships with a 262,144-token window natively, extensible to 1M via YaRN (with the usual short-sequence trade-off). It's also multimodal in — text, image and video input, text output — from a compact dense decoder.
Day-one local tooling
vLLM, SGLang, llama.cpp, LM Studio, Jan and Ollama all support it at launch, and Unsloth GGUF conversions are already up. Drop it in as a local coding or agent backend today, no API bill.