Qwen3 Coder Next

Qwen announced Qwen3-Coder-Next, an open-weight language model purpose-built for coding agents and local development. Built on Qwen3-Next-80B-A3B-Base, the model adopts a hybrid attention and Mixture-of-Experts architecture and is agentically trained at scale via executable task synthesis, environment interaction, and reinforcement learning to master long-horizon reasoning, tool usage, and failure recovery. The team reports over 70% on SWE-Bench Verified with the SWE-Agent scaffold and competitive results across SWE-Bench Multilingual, SWE-Bench Pro, TerminalBench 2.0, and Aider despite its small active footprint. The training recipe scales agentic signals rather than raw parameters, combining continued pretraining on code- and agent-centric data, supervised fine-tuning on high-quality agent trajectories, and domain-specialized expert training in software engineering, QA, and web/UX. Those experts are then distilled into a single deployment-ready model. The blog links to the full technical report, GitHub repository, Hugging Face, and ModelScope releases, emphasizing that the model matches or exceeds much larger open-source models on agent-centric evaluations at significantly lower inference cost.

alphaXiv's mirror of the Qwen3-Coder-Next technical report adds concrete benchmark context: the model scores 70.6% on SWE-Bench Verified using the SWE-Agent scaffold, narrowly beating DeepSeek-V3.2's 70.2% under identical evaluation conditions despite a much smaller 3B active footprint versus 37B. Reported runs allow up to 300 agent turns, showcasing the model's strength in long-horizon repository-level bug fixing where the agent must inspect code, call tools, and recover from execution failures. The report defines an executable task as a coding problem paired with a runnable code environment plus checks that distinguish a working fix from a failure, framing this as the foundation of the training approach. While Qwen3-Coder-Next nearly ties DeepSeek on SWE-Bench Verified, it trails competitors on several other plotted benchmarks, and the team releases base and instruction-tuned open-weight versions to support research and real-world coding agent development.

The Qwen team released Qwen3-Coder-Next on Hugging Face as an open-weight causal language model targeting coding agents, featuring 80B total parameters with only 3B activated per token. This efficient footprint delivers performance comparable to models with 10–20x more active parameters, making it cost-effective for always-on agent deployment. The model natively supports a 262,144-token context length and integrates with scaffolds including Claude Code, Qwen Code, Qoder, Kilo, Trae, and Cline. Architecture details include 48 layers with a hybrid layout of 12 repeating (3 Gated DeltaNet-MoE + 1 Gated Attention-MoE) blocks, 16 Gated Attention Q heads with 2 KV heads at head dimension 256, 32 Gated DeltaNet V heads, 512 experts with 10 activated plus 1 shared, and 2048 hidden dimension. The model operates exclusively in non-thinking mode without emitting blocks, and Qwen advises using the latest transformers version with reduced context length as a fallback for OOM issues on local hardware.

The Hugging Face model card for Qwen3-Coder-Next-Base is a first-party artifact from the Qwen team describing an open-weight language model purpose-built for coding agents and local development. It documents a Mixture-of-Experts architecture with 80B total parameters and 3B activated, organized as 48 layers using a hyb The card frames Qwen3-Coder-Next-Base as a pretraining-stage checkpoint intended to be a strong backbone for tool calling, scaffold/template adaptation, and error detection/recovery in coding agents, leaving headroom for downstream post-training. It notes that no inference provider currently hosts the Base weights on t