Qwen3-Coder-Next

// Overview

Qwen3-Coder-Next is an open-weight coding model from Alibaba's Qwen team, released on February 3, 2026 and built on the Qwen3-Next-80B-A3B-Base hybrid-attention architecture. It has 80 billion total parameters but activates only 3 billion per token, using a 512-expert mixture-of-experts router with 10 experts routed per token, and a 48-layer stack that combines Gated DeltaNet and Gated Attention. Alibaba positions it as a small but highly agentic coding model — performance comparable to dense models with 10–20× more active parameters, at sharply lower inference cost.

The model supports a 262,144-token native context window (256K) that can be extended to roughly one million tokens using YaRN. It was post-trained agentically at scale with large-scale executable task synthesis, environment interaction, and reinforcement learning, and is designed to drive coding agents through long-horizon tool-using workflows. Qwen3-Coder-Next operates in non-thinking mode only and does not emit <think> reasoning blocks. On Alibaba's reported benchmarks it scores 70.6% on SWE-bench Verified, 44.3% on SWE-bench Pro, and 36.2% on TerminalBench 2.0.

Qwen3-Coder-Next is released under the Apache 2.0 license with weights freely downloadable on Hugging Face and ModelScope (including FP8 and GGUF quantizations), and is hosted on Alibaba Cloud Model Studio as the `qwen3-coder-next` API model. It runs on Transformers, vLLM (≥0.15.0), SGLang (≥v0.5.8), Ollama, LM Studio, llama.cpp, MLX-LM, and KTransformers, with day-0 support on AMD Instinct GPUs, and exposes OpenAI-compatible API endpoints for drop-in agent integrations.

// Benchmarks

1. SWE-bench Verified70.6%
2. SWE-bench Pro44.3%
3. TerminalBench 2.036.2%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Very small active footprint — 3B activated of 80B total — makes inference cheap and CPU/edge-deployable, while matching the coding performance of much larger active-parameter models
• Hybrid Gated DeltaNet + Gated Attention architecture (from Qwen3-Next) gives long-context efficiency well beyond a standard transformer of the same active size
• 256K native context, extendable to 1M with YaRN — sufficient for repository-scale code understanding
• Open weights under Apache 2.0, with FP8 and GGUF quantizations published for low-VRAM hosting
• Built specifically for agentic coding via long-horizon reinforcement learning on executable tasks
• Wide runtime support: Transformers, vLLM, SGLang, Ollama, LM Studio, llama.cpp, MLX-LM, KTransformers, OpenAI-compatible endpoints

// Best for

• Self-hosted coding agents that need a small active-parameter footprint for cheap inference at scale
• Repository-scale code analysis and editing that exploits the 256K native (1M YaRN) context window
• Local development assistants on consumer GPUs using FP8 or GGUF quantizations
• Agentic CLI workflows that chain many tool calls and long-horizon reasoning over a codebase
• Drop-in replacement for closed coding APIs via OpenAI-compatible endpoints exposed by vLLM / SGLang
• Edge or CPU deployments where the 3B activated parameters fit memory budgets that 30B+ active dense models cannot

// How to access

// Qwen-Coder — every version

The full lineage of the Qwen-Coder line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.