Via alibaba-cloud.medium.com
Alibaba's new 27B-parameter model runs at 80 tokens per second on dual RTX 4090s, rivaling Anthropic's flagship on key benchmarks while fitting in 34GB of VRAM
A 27-billion-parameter model that matches or beats Anthropic’s Claude Opus 4.6 on coding benchmarks, runs on hardware you can buy at Micro Center, and ships under an open-source license. That’s the pitch behind Alibaba’s Qwen3.8-27B, and the early numbers are hard to argue with.
Community-quantized GGUF versions of the model are clocking roughly 80 tokens per second on two NVIDIA RTX 4090 GPUs, consuming just 34GB of total VRAM.
What Qwen3.8-27B actually delivers #
Released around August 14 by Alibaba’s Qwen team, the model is a dense architecture, meaning every parameter fires on every inference pass rather than routing through a mixture-of-experts system.
The native context window stretches to 262K tokens. With the integrated YaRN positional encoding system, that extends to 1 million tokens.
On SWE-bench Pro, a benchmark designed to test real-world software engineering ability, Qwen3.8-27B posted a score of 61.7. Anthropic’s Opus 4.6 Max scored 53.4 on the same benchmark.
The model also handles vision tasks, making it a multimodal system rather than a text-only engine. Alibaba’s self-reported benchmarks show competitive or superior performance across coding, agentic, and multimodal evaluations compared to Opus 4.6.
It’s licensed under Apache 2.0. Companies can modify it, deploy it commercially, and build products on top of it without licensing fees or restrictions.
The hardware story is the real headline #
Two RTX 4090s currently retail for somewhere around $3,200 to $4,000 total depending on the vendor and timing.
The 34GB VRAM footprint is notable because each RTX 4090 carries 24GB of VRAM, giving a dual-card setup 48GB total. That leaves meaningful headroom for system overhead, longer context windows, or running additional processes alongside the model.
The GGUF format itself deserves credit here. Community contributors have quantized the model for compatibility with popular inference frameworks including llama.cpp, vLLM, and Ollama.
Where Qwen3.8-27B fits in the broader landscape #
This release is part of a deliberate escalation from Alibaba’s Qwen team that began with the Qwen3 series launching in April and May of 2025. Earlier in August 2026, the team released Qwen3.8-Max, a larger model targeting cloud deployment. The 27B version is the local-friendly counterpart, optimized for accessibility.
The availability on Hugging Face and ModelScope ensures broad distribution, and the Apache 2.0 license removes the legal friction that has hampered adoption of some competing open models with more restrictive terms.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our