Qwen3-Coder
Qwen3-Coder is a family of open-weight large language models specialized for software engineering, developed by Alibaba's Qwen team (Tongyi Lab) and released under the Apache 2.0 license. The flagship release, Qwen3-Coder-480B-A35B-Instruct, was announced on July 22, 2025; it is a 480 billion parameter Mixture of Experts (MoE) model that activates 35 billion parameters per token, natively supports a 256,000 token context window (expandable to one million tokens with YaRN), and is positioned by Alibaba as its "most agentic code model to date" [1]. At launch it set the state of the art among open models on SWE-bench Verified with a score of 69.6 percent, performance Alibaba describes as "comparable to Claude Sonnet 4," which Alibaba presented as open-model parity with a frontier closed API on real-world repository tasks [1][2]. It ships alongside Qwen Code, an open-source command-line agent forked from Google's Gemini CLI, and supports 358 programming languages [1][17].
Qwen3-Coder is the second major coding-specialized release from the Qwen team, following Qwen2.5-Coder in late 2024. Where the earlier model focused on raw code completion and infill performance, Qwen3-Coder centers its design on agentic coding: long-horizon tasks where the model invokes tools, runs commands, edits files across a repository, and recovers from execution errors. Alibaba reports that the 480B variant achieves state-of-the-art results among open-weight models on agentic coding benchmarks, with the company claiming parity with Claude Sonnet 4 on several internal evaluations [1][5]. A more compact Qwen3-Coder-30B-A3B-Instruct followed on July 31, 2025, and a hybrid-attention successor, Qwen3-Coder-Next, was released in early 2026 with markedly higher SWE-Bench scores despite running only three billion active parameters [6][7].
Background and development
The Qwen team at Alibaba Cloud has shipped coding-specialized models since CodeQwen-7B in early 2024. The Qwen2.5-Coder line, released across multiple sizes from 0.5B to 32B in November 2024, drew attention for matching closed-source coding APIs on HumanEval and MBPP at the 32B size, and the 7B and 14B variants became common defaults for self-hosted coding assistants. With the broader Qwen3 family launching in April 2025 under a unified hybrid reasoning architecture, a coding-specialized successor was widely expected.
Qwen3-Coder was announced on the official @Alibaba_Qwen account on X on July 22, 2025 with the line: "Qwen3-Coder is here! We're releasing Qwen3-Coder-480B-A35B-Instruct, our most powerful open agentic code model to date." The release blog framed the model's purpose directly: "We're announcing Qwen3-Coder, our most agentic code model to date," available in multiple sizes but led by the 480B flagship [1]. The post emphasized two design choices that distinguish the model from earlier Qwen coding releases: native long-context training rather than position-encoding tricks at inference time, and a heavy investment in long-horizon reinforcement learning rather than supervised fine-tuning alone [1][5]. The release coincided with a wave of strong open-weight coding models that summer, including Moonshot's Kimi K2, Z.AI's GLM-4.5, and DeepSeek-V3.1, which together pushed open-source coding benchmarks within striking distance of frontier closed APIs.
The model family expanded over the following months. Qwen3-Coder-30B-A3B-Instruct, a smaller MoE variant aimed at single-GPU and consumer-hardware deployment, was published on Hugging Face on July 31, 2025 [19], and quantized GGUF, FP8, AWQ, and NVFP4 distributions appeared from community and vendor partners shortly after [3]. Qwen3-Coder-Next, an 80B total / 3B active hybrid-attention model built on the Qwen3-Next architecture, followed in early 2026 and became the high-water mark for the family's efficiency frontier [6][7].
What are the Qwen3-Coder variants?
Qwen3-Coder is published as several open-weight checkpoints with shared training recipes but different scales and architectures.
| Variant | Total parameters | Active parameters | Architecture | Native context | Release |
|---|---|---|---|---|---|
| Qwen3-Coder-480B-A35B-Instruct | 480B | 35B | MoE, 160 experts, 8 active | 256K (1M with YaRN) | July 22, 2025 |
| Qwen3-Coder-30B-A3B-Instruct | 30.5B | 3.3B | MoE | 256K | July 31, 2025 |
| Qwen3-Coder-480B-A35B-Instruct-FP8 | 480B | 35B | MoE, FP8 quantization | 256K | July 2025 |
| Qwen3-Coder-Next | 80B | 3B | Hybrid attention + sparse MoE, 512 experts, 10 active | 256K | February 3, 2026 |
The 480B model is the headline release and the only variant Alibaba evaluates against closed APIs in its launch materials. The 30B model is the practical default for local inference, since its 3.3 billion active parameters (30.5 billion total) make it light enough for single-GPU and consumer-hardware use [19]. Qwen3-Coder-Next is a distinct architectural lineage: it inherits the linear-attention plus gated-attention hybrid stack from Qwen3-Next-80B-A3B-Base and was post-trained explicitly for coding agents and local development [6].
All Qwen3-Coder checkpoints ship as Instruct variants only. Unlike the open Qwen3 base models, which expose a single weight set that toggles between thinking and non-thinking modes through a chat-template flag, Qwen3-Coder does not generate <think></think> blocks. The model card states that it supports only non-thinking mode [3].
Architecture
The flagship Qwen3-Coder-480B-A35B-Instruct uses a sparse MoE transformer with 62 decoder layers and grouped-query attention. Each multi-head attention block has 96 query heads and 8 key-value heads, and the MoE feed-forward layers contain 160 experts of which 8 are routed per token, giving the 35 billion active parameter count [3]. The model uses BF16 precision in its reference distribution, with FP8 and 4-bit quantizations published by Alibaba and community packagers for memory-constrained inference.
| Specification | Qwen3-Coder-480B-A35B |
|---|---|
| Total parameters | 480B |
| Active parameters per token | 35B |
| Decoder layers | 62 |
| Attention heads (query / key-value) | 96 / 8 |
| MoE experts | 160 |
| Active experts per token | 8 |
| Native context length | 262,144 tokens |
| Extended context (YaRN) | up to 1,048,576 tokens |
| Tokenizer | Qwen3 BPE (same as Qwen3 base) |
| Precision | BF16 (FP8 variant also released) |
| License | Apache 2.0 |
The attention pattern is dense across all layers; Qwen3-Coder does not use sliding-window or interleaved local-global attention for its primary checkpoint, which is one reason the model is heavy at inference time relative to its active parameter count. The successor Qwen3-Coder-Next switches to a hybrid linear-plus-gated-attention design that trades some attention precision for much lower KV cache requirements and what Alibaba claims is up to 10x higher throughput on repository-level tasks than dense models of similar total capacity [7].
How was Qwen3-Coder trained?
Alibaba describes the Qwen3-Coder training pipeline as a two-stage process: a coding-heavy pretraining phase and a multi-component post-training phase centered on reinforcement learning.
Pretraining
The pretraining corpus totals 7.5 trillion tokens, with code making up roughly 70 percent of the mix [1][5]. The remainder is general text and mathematics, included to preserve the natural-language reasoning and instruction-following capabilities the Qwen3 base models are known for. Alibaba reports using Qwen2.5-Coder as a quality filter and rewriter for noisy or low-quality code, an example of the synthetic data bootstrapping pattern that became common in 2025 frontier training. The Qwen3-Coder repository lists support for 358 coding languages [17].
Long-context capability is baked in during pretraining rather than added later. The native 256K context is achieved with extended-RoPE position encodings during pretraining, and the one-million-token figure is obtained by applying YaRN at inference time without further weight updates [3].
Post-training: Code RL and Agent RL
The post-training stage is where Qwen3-Coder diverges most sharply from earlier coding models. Alibaba describes two parallel reinforcement learning tracks:
The first is what the team calls Code RL: execution-driven reinforcement learning on a broad set of real-world coding tasks, including issue-fixing benchmarks, build-system tasks, and refactoring problems. Reward signals come from running the produced code against unit tests and reference solutions rather than from human preference labels. As Alibaba puts it, "Unlike the prevailing focus on competitive-level code generation in the community, we believe all code tasks are naturally well-suited for execution-driven large-scale reinforcement learning," which is why the team scaled Code RL across a broader set of real-world coding tasks, under the heading "Hard to Solve, Easy to Verify" [1].
The second is long-horizon RL, which the team calls Agent RL, where the model is trained to act as an autonomous coding agent across many turns. According to Alibaba, training was run in "20,000 independent environments in parallel" on Alibaba Cloud infrastructure, where the model used tools, browsed file systems, and interacted with mock browsers and shells to complete multi-step tasks [1][5]. This emphasis on agentic post-training is the main reason the model performs well on benchmarks like SWE-bench Verified, where the underlying task requires editing several files in a real repository and then passing a hidden test suite.
How does Qwen3-Coder score on benchmarks?
Alibaba's launch announcement claims that Qwen3-Coder-480B-A35B-Instruct sets new state-of-the-art results among open models on agentic coding, agentic browser use, and agentic tool use, and is comparable to Claude Sonnet 4 on real-world software engineering benchmarks [1]. The headline figure at launch was 69.6 percent on SWE-bench Verified, the best result reported for any open-weight model at the time, alongside 49.9 percent on WebArena for agentic browser use [1]. Independent reproductions and third-party leaderboards reported a range of SWE-bench Verified scores depending on the agent scaffold used.
| Benchmark | Qwen3-Coder-480B-A35B | Notes |
|---|---|---|
| SWE-bench Verified (official) | 69.6% | Alibaba launch figure, best open model at release |
| SWE-bench Verified (third-party) | 66.5% | Nebius OpenHands evaluation (Pass@1); results vary by scaffold |
| WebArena (agentic browser use) | 49.9% | Alibaba launch figure |
| SWE-Bench Pro | 38.7 | Hugging Face eval results [3] |
| Terminal-Bench 2.0 | 23.9 | Terminal-Bench leaderboard, Terminus 2 agent [3] |
| LiveCodeBench | 58.5% | Artificial Analysis, as listed on OpenRouter [10] |
The SWE-bench Verified score depends heavily on the agent harness used, since the benchmark measures end-to-end repository edits rather than single-shot code completion. In Nebius's own OpenHands evaluation, Qwen3-Coder-480B scored 66.5 percent Pass@1 [8]. By comparison, Claude Sonnet 4.5 was reported at about 77 percent and GPT-5 Codex at 74.5 percent on the same benchmark, each measured by its developer [15].
Qwen3-Coder-Next, despite running only 3 billion active parameters, is reported to reach 70.6 percent on SWE-bench Verified with SWE-Agent, 71.1 percent with MiniSWE-Agent, and 71.3 percent with OpenHands, matching or exceeding the original 480B flagship on the same scaffolds [6]. This is the headline efficiency claim of the Qwen3-Coder line: that an 80B / 3B-active hybrid model can match or beat a 480B / 35B-active dense-attention MoE on the most-watched coding benchmark.
What is the Qwen Code CLI?
Alongside the model weights, Alibaba released Qwen Code, an open-source command-line agent designed to run Qwen3-Coder locally or against a remote API. The tool is adapted from Google's Gemini CLI codebase under its Apache 2.0 license, with "customized prompts and function calling protocols" reworked to match Qwen3-Coder's tool-call schema and the parser tightened around Qwen's structured output [1][4][9]. The CLI exposes the model as an interactive shell assistant that can read and modify files in the working directory, run commands, and chain tool calls into multi-step workflows.
Qwen Code targets the same developer niche as Claude Code, GitHub Copilot CLI, and OpenAI's Codex CLI, but its open-source license and compatibility with self-hosted Qwen3-Coder endpoints make it the default agent shell for users who want to run a coding assistant against on-premise hardware. Integrations exist for VS Code, Zed, and JetBrains IDEs through IDE plugins documented in the Qwen Code repository, and the CLI can also be pointed at OpenRouter, Alibaba Cloud's Model Studio API, or any OpenAI-compatible endpoint serving Qwen3-Coder weights [4].
The Qwen team also published explicit support for popular third-party coding agents. The launch blog gives setup instructions for Claude Code (through an Alibaba Cloud proxy API) and Cline, and the model card describes "a specially designed function call format" for platforms such as Qwen Code and CLINE [1][3].
How does Qwen3-Coder compare with frontier coding models?
Qwen3-Coder is most often compared with Claude Sonnet 4, GPT-5 Codex, and other open coding models such as Kimi K2 and GLM-4.5. The picture that emerged across third-party evaluations during late 2025 and into 2026 is one where Qwen3-Coder consistently lands in the second tier of coding capability behind the leading closed APIs, but distinguishes itself on cost, openness, and local-inference viability.
| Model | SWE-bench Verified | Context | License | Approx. input cost per million tokens |
|---|---|---|---|---|
| Claude Sonnet 4.5 | ~77% | 200K | Closed | $3.00 |
| GPT-5 Codex | 74.5% | 400K | Closed | $1.25 |
| Qwen3-Coder-480B-A35B | ~67 to 70% | 256K (1M extrap) | Apache 2.0 | $0.30 (OpenRouter, Nov 2026) |
| Qwen3-Coder-Next | ~71% | 256K | Apache 2.0 | $0.12 (OpenRouter, Nov 2026) |
| DeepSeek-V3.2 | ~70% | 128K | MIT | $0.26 (OpenRouter, Nov 2026) |
| GLM-4.7 | ~74% | 200K | MIT | $0.60 (OpenRouter, Nov 2026) |
The cost gap is the easiest pitch for Qwen3-Coder. At a fraction of the per-token cost of closed coding APIs (as of November 2026, OpenRouter lists the 480B model at $0.30 per million input tokens, against $1.75 for GPT-5.3 Codex), the model is a common choice for high-volume agentic workflows where each task may consume hundreds of thousands of tokens [10][16]. The SWE-bench figures in the table come from each developer's or Qwen's own reporting [1][15][18]. Practical reviews suggest that Claude Sonnet 4 and GPT-5 Codex still produce more complete and reliable implementations on complex refactors, while Qwen3-Coder is competitive on isolated function-level tasks and noticeably faster per token [11][12].
Within the open-weight coding tier, the comparisons are tighter. Kimi K2 from Moonshot AI, released in mid-2025, also targets agentic coding at a similar scale, and the two models trade wins on different SWE-bench scaffolds. GLM-4.5 from Z.AI scored lower on SWE-bench Verified (64.2 percent in its technical report), while the later GLM-4.7 scores higher (74.2 percent with SWE-Agent in Qwen's comparison) [18][21], but Qwen3-Coder's tooling ecosystem, especially the Qwen Code CLI and the strong third-party agent integrations, has made it one of the most widely used open coding models.
Qwen3-Coder-Next
Qwen3-Coder-Next is a successor model released in early 2026 that pushes the family further toward efficient sparse architectures. It uses the Qwen3-Next-80B-A3B base, a hybrid model that combines linear attention (Gated DeltaNet) with gated softmax attention in a repeating three-to-one layer pattern [20] and replaces the 160-expert MoE of the 480B flagship with a 512-expert sparse network that routes 10 experts per token. The result is 80 billion total parameters with only 3 billion activated per forward pass [6][7].
Winbuzzer's coverage of the release reported a 70.6 percent SWE-bench Verified score, just ahead of DeepSeek-V3.2 (70.2 percent) and behind GLM-4.7 (74.2 percent), and said the model outperformed Claude Opus 4.5 on the SecCodeBench secure-code-generation benchmark [7]. Alibaba's benchmark table for the model claims SWE-bench Verified scores between 70.6 and 71.3 percent depending on agent harness, SWE-Bench Pro of about 44.3 percent, and competitive numbers on SWE-Bench Multilingual, Terminal-Bench 2.0, and Aider edit benchmarks [6].
Qwen3-Coder-Next is also notable for its emphasis on adaptability to different agent scaffolds. The Qwen team's technical report says models trained on a single tool chat template tend to overfit to it, and that because "real-world agent systems use a wide variety of formatting conventions," Qwen3-Coder-Next was trained on diverse tool chat templates and formats [18].
Is Qwen3-Coder open source?
Qwen3-Coder is published under the Apache 2.0 license, which permits commercial use, modification, and redistribution without copyleft requirements [3]. Weights for the 480B, 30B, and Next variants are hosted on Hugging Face under the official Qwen organization, with quantized distributions in GGUF, FP8, AWQ, GPTQ Int4, and NVFP4 formats available from community and vendor accounts. The reference code repository is QwenLM/Qwen3-Coder on GitHub, and the Qwen Code CLI is hosted at QwenLM/qwen-code [1][4].
For users who do not want to run the weights themselves, Qwen3-Coder is accessible through Qwen Chat (Alibaba's web UI), Alibaba Cloud's Model Studio API, and a long list of third-party inference providers including OpenRouter, Together AI, Fireworks AI, DeepInfra, and Hyperbolic. As of November 2026, OpenRouter lists the 480B model at $0.30 per million input tokens and $1 per million output tokens and no longer lists a free variant [10][16]. Alibaba also sells API access directly through Model Studio [2].
Reception
Qwen3-Coder was generally well received in the open-source AI community on release. Hacker News commenters highlighted the model as the first open-weight coder to feel competitive with Claude on real agentic workflows, and the immediate availability of a usable CLI agent at launch was widely noted as a differentiator from earlier open releases that shipped weights without first-party tooling. VentureBeat's launch coverage led with the framing that Qwen3-Coder "might be the best coding model yet," and AI press more broadly treated the release as an escalation in the open-versus-closed coding model competition, with VentureBeat and MarkTechPost both running prominent stories on the 480B flagship and the later Qwen3-Coder-Next [7][13][14].
Critical responses focused on three main issues. First, the 480B model is expensive to host: even with sparse activation, it requires several hundred gigabytes of memory in BF16 and is impractical for individual developers to self-host without aggressive quantization. Second, despite Alibaba's parity claims, side-by-side comparisons with Claude Sonnet 4 on complex multi-file refactors generally favored Claude for reliability, even when Qwen3-Coder produced acceptable code more often than baseline open models. Third, the Qwen Code CLI's adaptation from Gemini CLI inherited some of the rough edges of the original tool, and users reported the experience was less polished than Claude Code at launch.
The release of Qwen3-Coder-30B and later Qwen3-Coder-Next addressed the first issue directly. The 30B model fit on consumer hardware, and the 80B / 3B-active Next variant fit on workstation-class machines while approaching the benchmark performance of the flagship, which prompted a second wave of community deployment.
Continue reading
- Qwen3Qwen3 is the third-generation family of large language models developed by the Qwen Team at Alibaba Cloud (also known as Tongyi Qianwen lab).Mentioned 69 times above
- QwenQwen is a family of large language models and multimodal models developed by the Qwen Team at Alibaba Cloud.Mentioned 23 times above
- SWE-benchSWE-bench is an execution-based benchmark for evaluating whether a language-model system can resolve real software issues.Mentioned 18 times above
- Qwen2.5-CoderQwen2.5-Coder is the code-specialized series within the Qwen2.5 generation of large language models developed by the Qwen team at Alibaba.Mentioned 3 times above
- Coding agentA coding agent is a software system that uses a large language model to carry out programming work by acting on a development environment rather than only producing text for a person to copy.Mentioned 3 times above
- Claude CodeClaude Code is an agentic software development product made by Anthropic. It uses models from the Claude family to inspect software projects, propose and apply changes, run development tools, and report…Mentioned 3 times above
More from See also
More to read
Topics:Chinese AIDeveloper ToolsLarge Language ModelsOpen Source AI
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Qwen Team. "Qwen3-Coder: Agentic Coding in the World." Qwen Blog, July 22, 2025. qwenlm.github.io/...qwen3-coder
- ^1 ^2Alibaba Cloud. "Alibaba Unveils Cutting-Edge AI Coding Model Qwen3-Coder." Alibaba Cloud Press Room, July 2025. alibabacloud.com/...ing-edge-ai-coding-model-qwen3
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Qwen Team. "Qwen3-Coder-480B-A35B-Instruct." Hugging Face model card. huggingface.co/...Qwen3-Coder-480B-A35B-Instruct
- ^1 ^2 ^3QwenLM. "qwen-code: An open-source AI agent that lives in your terminal." GitHub repository. github.com/...qwen-code
- ^1 ^2 ^3 ^4Data Science Dojo. "Qwen3 Coder: The Open-Source AI Coding Model Redefining Code Generation." 2025. datasciencedojo.com/...qwen3-coder
- ^1 ^2 ^3 ^4 ^5 ^6Qwen Team. "Qwen3-Coder-Next: Pushing Small Hybrid Models." Qwen Blog. qwen.ai/blog
- ^1 ^2 ^3 ^4 ^5 ^6Winbuzzer. "Alibaba's Qwen3-Coder-Next Activates Just 3B of 80B Parameters For Improved Efficiency." February 4, 2026. winbuzzer.com/...ce-sparse-moe-coding-model-xcxwbn
- ^Nebius. "OpenHands trajectories with Qwen3-Coder-480B-A35B-Instruct." Nebius Blog. nebius.com/...s-trajectories-with-qwen3-coder-480b
- ^DataCamp. "Qwen Code CLI: A Guide With Examples." datacamp.com/...qwen-code
- ^1 ^2 ^3OpenRouter. "Qwen3 Coder 480B A35B - API Pricing & Benchmarks." openrouter.ai/...qwen3-coder
- ^Daily Dose of DS. "Compare Qwen 3 Coder vs. Sonnet 4 for Code Generation." blog.dailydoseofds.com/...qwen-3-coder-vs-sonnet-4
- ^LLM Stats. "GPT-5.3 Codex vs Qwen3-Coder Comparison." llm-stats.com/...gpt-5.3-codex-vs-qwen3-coder
- ^MarkTechPost. "Qwen Team Releases Qwen3-Coder-Next: An Open-Weight Language Model Designed Specifically for Coding Agents and Local Development." February 3, 2026. marktechpost.com/...g-agents-and-local-development
- ^VentureBeat. "Qwen3-Coder-480B-A35B-Instruct launches and it 'might be the best coding model yet'." July 2025. venturebeat.com/...ht-be-the-best-coding-model-yet
- ^1 ^2TechRadar. "OpenAI launches GPT-5-Codex with a 74.5% success rate on real-world coding." September 2025. techradar.com/...success-rate-on-real-world-coding
- ^1 ^2OpenRouter. "Models API" (model list with per-token prices, retrieved November 2026). openrouter.ai/...models
- ^1 ^2QwenLM. "Qwen3-Coder." GitHub repository README. github.com/...Qwen3-Coder
- ^1 ^2 ^3Qwen Team. "Qwen3-Coder-Next Technical Report." arXiv:2603.00729, February 2026. arxiv.org/...2603.00729
- ^1 ^2Qwen Team. "Qwen3-Coder-30B-A3B-Instruct." Hugging Face model card. huggingface.co/...Qwen3-Coder-30B-A3B-Instruct
- ^Qwen Team. "Qwen3-Coder-Next." Hugging Face model card. huggingface.co/...Qwen3-Coder-Next
- ^GLM-4.5 Team. "GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models." arXiv:2508.06471, August 2025. arxiv.org/...2508.06471
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
6 revisions · v6 · 3,302 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check of the full article against its cited sources (95 claims, 20 sources fetched); 7 material + 28 minor fixed
Cite this page: AI Wiki. "Qwen3-Coder." aiwiki.ai, updated 28 Oct 2026, fact-checked 5 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/qwen_3_coder

