Qwen3.8-27B: Apache-2.0 coding near Opus, runs on a 4090
The dense, vision-capable 27B hits 61.7 on SWE-bench Pro with a 262K context, and 4-bit GGUFs are live for llama.cpp — its 2.4T sibling needs a rack.
Copy markdown
It edges Opus 4.6 on SWE-bench Pro
Alibaba's own table puts the 27B at 61.7 on SWE-bench Pro against a cited 53.4 for Opus 4.6 Max, plus 79.0 on QwenSWEBench and 89.2 on GPQA Diamond. Treat same-lab numbers and differing agent harnesses with caution — but a 27B trading blows with a frontier closed model is the story.
Runs on a single 24GB card tonight
The 4-bit GGUF is 14–17GB, so it fits an RTX 4090; FP8 wants ~28GB, BF16 ~56GB. Pull it with `ollama run qwen3.8:27b`, or `vllm serve Qwen/Qwen3.8-27B` for an OpenAI-compatible endpoint. Throughput lands around 10–160 tok/s depending on quant.
Apache 2.0 — ship it commercially, no strings
The 27B is plain Apache 2.0: fine-tune it, bake it into a product, and sell that product with no revenue share or usage gate. That's the opposite of the custom, revenue-share license Alibaba put on the 2.4T Qwen3.8-Max.
Multimodal, with a 262K context
It ships a vision encoder and a 262,144-token native window, extendable toward 1M via YaRN — a local model you can point at whole repos and screenshots, not just chat. Budget extra VRAM, since the KV cache grows with that long context.
Elsewhere: the 2.4T sibling needs a datacenter
Qwen also opened Qwen3.8-Max, a 2.4T-parameter MoE with 95B active — but it's text-only, custom-licensed, and wants ~5TB of storage or a 72-GPU rack to serve. Impressive, but the 27B is the one you'll actually run.