Two weeks after the 2.4T MoE flagship, Alibaba dropped the weights for its dense little brother. Qwen3.8-27B is a 27B dense model under Apache 2.0, and at 4-bit it needs about 17GB — one 3090, one 4090, or a Mac with enough unified memory. Unsloth had GGUF quants up within hours of the FP8 release. Two HN threads on the same launch cleared 200 combined points in a day.
What it actually does
It’s a general model built for agentic coding: 61.7 on SWE-bench Pro, 90.3 on LiveCodeBench v6, 73.0 on Terminal-Bench 2.1. Native 262K context, extensible to 1M with YaRN. Reasoning is on by default with a reasoning_effort dial, and it takes images and video too. The predecessor Qwen3.6-27B was already the local crowd’s default; this replaces it.
Running it
llama.cpp, vLLM, SGLang, and Ollama all load it today. Don’t want the hardware? Alibaba serves it through Model Studio, so the same weights work as a hosted API — build the agent locally, ship it on the cloud endpoint, no rewrite.
The interesting part isn’t the scores. It’s that a model this close to frontier coding performance now runs on a gaming GPU with no API bill attached.
You Might Also Like
- Qwen3 8 max Goes ga Alibabas 2 4t Model Beats Claude Fable 5 on Terminal Bench Weights Open Next Week
- Alibabas Qwen 3 6 27b Scores 77 2 on swe Bench and Runs on one Consumer gpu
- Hugging Face Speech to Speech Open Source Local Voice Agents is the Openai Realtime Clone you can run on Your own gpu
- Figure 03 Helix 02 Runs 67 Hours Autonomous With one Error 2026 Production Already Sold out
- 1700 hn Points in one day Chinas Open Weights ai Strategy is Winning Werd io and Stratechery Open Fire Together
Top AI Product 
Leave a comment