Alibaba Open-Sources Qwen3.8-27B: Runs on One GPU, Codes Near Claude Opus Level
rohanpaul_ai · x · 2026-08-15
Alibaba has released the weights for Qwen3.8-27B: a 27B-parameter native multimodal dense model built for local deployment under Apache 2.0, outperforming Qwen3.7-Plus overall.
Key specs:
- 262K native context, extendable to 1M tokens via YaRN; reasoning can be disabled or adjusted per request
- Supports Transformers, vLLM, SGLang and local quantizations, putting unusually capable multimodal agent work within single-machine range
- AMD's initial testing: up to 51.8 tokens/sec on a single Radeon AI PRO R9700 (24GB VRAM)
Coding performance lands close to, and sometimes above, Claude Opus 4.6 Max: SWE-bench Pro 61.7 vs 53.4, CoWorkBench 70.7 vs 68.2, and OSWorld 84.3 vs 72.7. A 27B local model that is legitimately frontier-class on several coding/agent benchmarks. The open weights for Qwen3.8-2.4T-A95B (Max-level) were also released recently.
Related event: Alibaba Open-Sources Qwen3.8-27B Multimodal Model That Runs on a Single GPU(2 posts)→
More from Models
- Frontier Model Releases Rise, but Sub-Frontier Releases Are Soaring — AaronBergman18 · 2026-08-15
- Cost-efficient SFT trick: "loss to zero" validates data value — tokenbender · 2026-08-15
- Local Qwen3.8-27B One-Shots Super Mario Clone: User Impressed — MikeNonect · 2026-08-15
- Uncensored Qwen3.8-27B GGUF Model Trends on Hugging Face — JonathanColetti · 2026-08-15
- Vision Model Challenge: Meter Reading Should Be 47461, Can Your Model Get It? — MrMrsPotts · 2026-08-15
- Test Shows Qwen3.8 Offers More Mature Security Judgment than 3.6 — thomble · 2026-08-15