Qwen3-Coder-Next

2026-02-03 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba (Qwen team) has released Qwen3-Coder-Next, an open weights coding model of 80 billion parameters with 3 billion active, designed for coding agents, under Apache 2.0.

### Context The coding models of the previous generation (Qwen3-Coder family) relied on large MoEs with classic attention. In early 2026, the coding agent market (Claude Code, Qwen Code, Cline, etc.) calls for models able to run at reduced cost on entire repositories. Qwen3-Coder-Next takes the experimental Qwen3-Next architecture (hybrid linear attention) and steers it toward agentic use. ### What the model brings - Active parameters / capacity ratio: 80B in total for 3B activated per token, which aims at performance comparable to that of models with 10 to 20 times more active parameters (vendor claim). - Agentic training: according to the model card, a dedicated recipe (synthesis of executable tasks, interaction with environments, reinforcement learning) targets long reasoning, tool use and recovery after execution failure. - IDE / CLI integration: 256k tokens of context and adaptation to several tool templates (Claude Code, Qwen Code, Qoder, Kilo, Trae, Cline cited by the vendor). - Non-thinking mode only: the model does not produce reasoning blocks (`<think>`), which simplifies use in an agent loop. - A technical report followed on arXiv (2603.00729, submitted on February 28, 2026), with the Base and instruct versions in open weights. ### Architecture and specifications - Parameters: 80B in total, 3B activated (79B excluding embeddings); 48 layers; layout 12 x (3 x (Gated DeltaNet then MoE) then 1 x (Gated Attention then MoE)). - MoE: 512 experts, 10 activated plus 1 shared, expert intermediate dimension of 512. - Context: 262,144 tokens natively. Text only. - License: Apache 2.0 (free commercial use, no user threshold). - Weights: Hugging Face, Qwen/Qwen3-Coder-Next (Base, FP8 and GGUF versions published). Knowledge cutoff: not disclosed. ### Cost and access | Item | Value | |---|---| | Official API pricing (input / output, USD per million tokens) | Not consulted at the primary source, not disclosed here | | Self-hosting | vLLM 0.15.0 or later; the model card's example launches the service with tensor parallelism on 2 GPUs with 256k context | | Minimum hardware | Not disclosed by the vendor | ### Benchmarks The model card's figures are provided as images (SWE-bench Pro, comparison chart) and not as a text table. I therefore could not verify them line by line at the primary source, and no comparative table with closed models is reproduced here. Third-party sites cite about 74% on SWE-bench Verified, a figure not verified at the primary source. ### Versus closed models The vendor presents the model as comparable to models with 10 to 20 times more active parameters, without the table being reproducible here. Lacking a verified numerical comparison with a contemporary closed model, no conclusion on the gap is put forward. What is established: the Apache 2.0 license, the 3B active parameters and the local usage mode make the model self-hostable, which meets sovereignty and cost requirements that closed APIs do not cover. The choice between open and closed here comes down to control over deployment more than to a demonstrated score gap. ### Key takeaways Qwen3-Coder-Next sets Qwen's direction for 2026: open models that are small in active parameters, built for agentic use. Its strength for the enterprise lies in self-hosting a coding assistant under Apache 2.0.

Tags : LLM, Open Source, Alibaba, Qwen, Apache 2.0, MoE, Code