Qwen3-Coder-Next is an open-weight coding-focused large language model released by Alibaba's Qwen team in February 2026 as part of the Qwen family. It succeeds the 2025 Qwen3-Coder line and is built on the Qwen3-Next-80B-A3B base model, an ultra-sparse design with about 80 billion total parameters of which roughly 3 billion are active per token (developer-reported).1 In some community reporting the release has also been referenced under the informal label "Qwen 3.6."

Design and training

Unlike the dense and conventional mixture-of-experts Transformer variants used in earlier Qwen3 models, Qwen3-Coder-Next inherits the hybrid architecture of the Qwen3-Next base: blocks of Gated DeltaNet linear-attention layers interleaved with gated standard attention, each followed by sparse expert layers. Third-party technical write-ups describe 512 experts with about 10 routed per token and native support for a 256,000-token context window.2 The Qwen team states that the model was post-trained specifically for coding agents through large-scale executable task synthesis, interaction with software environments, and reinforcement learning, rather than through static instruction tuning alone.1

Performance and positioning

Developer-reported results, as of February 2026, include scores above 70 percent on SWE-Bench Verified when the model is run inside the SWE-Agent scaffold, placing it near much larger proprietary systems on agentic software-engineering tasks despite its small active-parameter budget.1 Press coverage also reported that on SecCodeBench, a benchmark for security-oriented code generation and vulnerability repair, the model outperformed Claude Opus 4.5 in code-generation scenarios (61.2 versus 52.5 percent, as reported).3 The sparse design is aimed at high-throughput inference and local deployment, a niche also contested by open coding models such as DeepSeek-Coder-V2, Codestral, Devstral, and Code Llama, as well as open-weight generalist agents like Kimi K2 and GLM-4.6.

Availability

The weights were published under the Apache 2.0 license, permitting commercial use, with checkpoints distributed through Hugging Face and ModelScope and hosted access offered via Alibaba Cloud. Community quantizations for consumer hardware appeared shortly after release, reflecting the model's positioning as a locally runnable alternative to proprietary coding assistants such as those in the GPT-5.1 and Claude lines.2

References


  1. Qwen team, "Qwen3-Coder-Next: Pushing Small Hybrid Models," February 2026. 

  2. Community technical guide, "Qwen3-Coder-Next: The Complete 2026 Guide to Running Powerful AI Coding Agents Locally," DEV Community, 2026. 

  3. VentureBeat, "Qwen3-Coder-Next offers vibe coders a powerful open source, ultra-sparse model with 10x higher throughput for repo tasks," 2026.