1. Summary

Qwen3-Coder-Next is a causal language model based on open-weight developed by the Qwen team, specifically designed for coding agents and local development environments. It has 3B active parameters. This allows it to achieve performance comparable to models with 10 to 20 times more active parameters at a significantly lower cost.

Additionally, agent capabilities such as long-term inference, complex tool use, and recovery from execution failures have been enhanced. With a native context length of 256k tokens, it can be integrated into various development environments such as Claude Code and Cline. Note that this model only supports non-thinking mode and does not generate thinking blocks.

2. Bunrin Bench (BUNRIN LABO Original Test)

Results of the "Bunrin Bench" machine-scoring benchmark (covering 7 categories including output contracts, fact retention, instruction following, and code traps) conducted on local hardware by we. For the methodology, refer to About Bunrin Bench.

Target Test Version Total Contract Number Constraint Hearsay Code Non-Hallucination Title Speed Test Date
qwen3-coder-next-4.5bit (MacBook Pro (M5 Max, 128GB)) ijiwaru-v1 40/41 (98%) 3/3 9/9 6/6 2/3 11/11 5/5 4/4 58.7 tok/s 2026-09-01

Measurement Environment:

  • qwen3-coder-next-4.5bit — Model: spicyneuron/Qwen3-Next-Coder-MLX-4.5bit (MLX version). Runtime: oMLX 0.6.3 (mlx 0.32.0, mlx-lm 0.31.3). Sampling: temperature 1 / top_p 0.95 / top_k 40

3. Various Benchmarks

The Hugging Face Model Card includes benchmark comparison charts and SWE-bench Pro results as images. The official blog shows performance across SWE-Bench, TerminalBench 2.0, and Aider benchmarks, reporting a trend where performance improves by increasing the number of agent turns, particularly in SWE-Bench Pro.

Since specific numerical values were not provided in the text materials, it is necessary to refer to the images in the model card or the official blog to verify the numbers.

4. Official announcements

  • On February 4, 2026, the Qwen team announced the open-weight release of Qwen3-Coder-Next, based on information from the Hugging Face Model Card.
  • The technical report (Qwen3-Coder-Next Technical Report) has been published on GitHub.
  • The official blog provides details on the model architecture, agent learning methods, and benchmark evaluations.
  • For deployment, sglang v0.5.8 or higher and vLLM 0.15.0 or higher are recommended. For local use, support is reported for Ollama, LMStudio, MLX-LM, llama.cpp, and KTransformers.

5. Real-world performance (Community reception)

On February 4, 2026, the announcement of this model was posted on Hacker News, attracting significant attention with 735 points and 429 comments (HN Post). The official blog explains a design philosophy focused on scaling agent learning signals.

Training via large-scale executable task synthesis and reinforcement learning is emphasized. Sufficient information regarding specific user experiences from the community has not yet been gathered.

Meanwhile, on Reddit, there is a discussion regarding the possibility that frontier models may sabotage local AI implementations (Reddit). One user reported that frontier models exhibited behavior contrary to instructions.

That user also mentioned examples where frontier models recommended Qwen3-Coder-Next (which is an older model).

6. Recommended parameters

The recommended sampling parameters in the "Best Practices" section of the Hugging Face Model Card are as follows:

Parameter Recommended Value
temperature 1.0
top_p 0.95
top_k 40

Source: Hugging Face Model Card "Best Practices" section.

7. Sources