Qwen 3.6-27B tops Devstral for local coding crown

// 46d agoMODEL RELEASE

Qwen 3.6-27B tops Devstral for local coding crown

Alibaba's newly released Qwen 3.6-27B model is challenging Mistral’s Devstral Small 2 for the title of best local coding LLM on 24GB GPUs. Both models represent the pinnacle of "agentic" software engineering performance for developers running consumer hardware like the RTX 3090.

// ANALYSIS

Qwen 3.6-27B represents a generational shift in local coding performance, making it the clear choice over Devstral for most 24GB VRAM users.

–Qwen’s 77.2% SWE-bench Verified score significantly outpaces Devstral’s 68.0%, rivaling closed-source giants like Claude 4.5.
–The hybrid Gated DeltaNet architecture provides native long-context stability up to 262k tokens, essential for repository-wide refactoring.
–While Devstral remains highly efficient for simple terminal tasks, Qwen’s multimodal "Thinking Mode" offers superior reasoning for complex frontend and UI/UX engineering.
–Both models fit comfortably on a single 3090 using FP8 or Q6 quantizations, though Qwen's newer architecture yields faster prompt processing.

// TAGS

qwen-3.6devstralmistralllmai-codingagentopen-weightsgpu

DISCOVERED

46d ago

2026-04-28

PUBLISHED

46d ago

2026-04-28

RELEVANCE

9/ 10

AUTHOR

szansky

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL48m ago

MiniMax highlights the open-weights availability of its M3 model on Hugging Face following the suspension of Anthropic's closed models.

Following the suspension of Anthropic's closed models, rival lab MiniMax highlighted the availability of its open-weight M3 model on Hugging Face. The move positions MiniMax-M3 as a reliable open alternative for developers looking to mitigate the risks associated with proprietary API dependencies.

RESEARCH1h ago

Hexo Labs introduces SIA (Self-Improving AI), a framework that treats agents as editable systems by co-optimizing their scaffolds and internal weights in a unified loop.

SIA is an open-source framework developed by Hexo Labs that allows an AI meta-agent loop to autonomously edit and improve all facets of a target agent, including its external scaffold (such as tools, parsers, and verifiers) and its model weights. By treating both code and weights as optimization targets, SIA achieves significant performance gains across tasks, including 70.1% accuracy on the LawBench legal benchmark, optimization of Triton CUDA kernels down to 1,017 µs, and a 0.289 mse_norm reduction in scientific denoising tasks.

POLICY1h ago

Amazon report triggers Claude Fable 5 ban

Amazon researchers discovered a critical security vulnerability in Anthropic's Claude Fable 5, leading CEO Andy Jassy to report the issue directly to the U.S. government. In response, the Department of Commerce imposed emergency export controls, prompting Anthropic to disable global access to both its Fable 5 and Mythos 5 models.

Qwen 3.6-27B tops Devstral for local coding crown