Qwen3-Coder-480B-A35B-Instruct is the May 2025 specialized coding model from Alibaba Qwen, a 480-billion-parameter mixture-of-experts language model with 35 billion active parameters per token, 256K-token native context (extensible to one million tokens via YaRN), and Apache 2.0 open-weights licensing. The model is positioned specifically for agentic coding workflows, repository-scale code understanding, and browser-use automation, with first-class integration into Alibaba's Qwen Code CLI and the CLINE agentic-coding ecosystem. As of May 2026, the Coder line remains a principal entry in the open-weights coding-specialist category alongside DeepSeek Coder and the broader Qwen 3 series, and is hosted across the major third-party inference providers including Fireworks AI.
At a glance
- Lab: Alibaba Qwen.
- Released: May 14, 2025 (corresponding technical paper arxiv 2505.09388).
- Modality: Text generation, specialized for code. No native image, video, or audio support.
- Open weights: Yes. Apache 2.0 license, including commercial use.
- Architecture: Sparse mixture-of-experts. 480 billion total parameters with 35 billion active per token. 62 layers, 96 query attention heads, 8 key-value heads, 160 total experts with 8 activated per forward pass.
- Context window: 262,144 tokens (256K) natively, extendable to approximately 1 million tokens via YaRN positional scaling.
- Pricing: Open weights, free to self-host. Hosted-inference pricing through third-party providers (including Fireworks AI, Novita, and others) varies; the model is positioned at materially lower per-token cost than the closed-source coding-frontier alternatives.
- Distribution channels: Hugging Face Hub, Ollama, LMStudio, MLX-LM, llama.cpp, KTransformers, vLLM, SGLang, HuggingChat inference provider, Docker Model Runner, the Qwen Code command-line tool, and CLINE agentic-coding integration.
Origins
Qwen3-Coder-480B-A35B-Instruct is the largest specialized coding model in Alibaba Qwen's open-weights catalog as of mid-2025 and the principal entry in the dedicated coding-model branch of the Qwen 3 generation. Alibaba Qwen's coding-model strategy has consistently maintained a dedicated coding line alongside the general-purpose Qwen models, dating to the Qwen Coder 7B and CodeQwen 1.5 releases in 2024 and continuing through the Qwen 2.5 Coder 32B release. The Qwen 3 Coder 480B-A35B Instruct release scales the dedicated coding line to the same parameter-count tier as the general-purpose Qwen 3 235B-A22B and Qwen 3 32B variants.
The architectural choice of 35 billion active parameters places Qwen 3 Coder between two distinct design philosophies in the open-weights coding category. Smaller coding specialists (the Qwen 2.5 Coder 32B and similar) operate at single-digit billion active parameters and are optimised for IDE-embedded completion latency. Larger generalist models that happen to be strong at code (DeepSeek V4 Pro, Qwen 3 235B-A22B) operate at 22-49 billion active parameters but are not specifically tuned for code. Qwen 3 Coder 480B-A35B sits at the larger active-parameter end of the coding-specialist band, trading deployment cost for agentic-task capability ceiling.
The release form is "Instruct" rather than "Thinking": the model operates without <think> blocks, in contrast to the reasoning-tuned models (Qwen 3, QwQ 32B, the Qwen 3.6 family) that emit thinking traces by default. The non-thinking instruction-tuned mode is positioned as the appropriate default for IDE-integrated completion and tool-call workflows where latency matters and the explicit reasoning trace would be a UI-and-latency drag.
The integration into Qwen Code (the Alibaba-maintained command-line coding agent) and CLINE (the open-source agentic-coding IDE extension) provides the principal application-layer surface for the model. Both integrations expose the agentic-coding capability through standardized tool-call interfaces (file editing, shell commands, browser automation) rather than through raw token-completion APIs.
Capabilities
The Qwen 3 Coder capability profile is concentrated on agentic coding, repository-scale code understanding, and browser-use automation.
Agentic coding is the headline positioning. The model is designed to operate against a tool-call interface (file editing, shell commands, version-control operations, search) rather than to generate code in isolation. The training approach explicitly targets the multi-turn agent loop that integrated development environments and command-line agents use: parse the user request, plan the multi-step approach, execute tool calls, evaluate the results, iterate.
Repository-scale code understanding is enabled by the 256K-token native context window. At standard tokenisation rates, 256K tokens accommodates approximately one million lines of code, which is sufficient for most monorepos at the single-package level and for many multi-package repositories at the read-only-analysis level. The optional 1-million-token YaRN-extended context further extends the addressable scope to multi-repository codebases.
Browser-use automation is the third principal capability. The model is positioned for browser-agent workflows including web scraping, form completion, multi-page navigation, and search-driven information gathering. The Evasion Bench score of 78.16 reflects this positioning: the benchmark evaluates browser-based agent capability and the high score places the model in the leading group on this specific axis.
Tool-call format uses a specialized function-call structure documented in the model card, similar to but not identical to the OpenAI function-call schema. The custom format is the structural mechanism that the Qwen Code and CLINE integrations use; developers adapting the model to other agent frameworks need to map the format appropriately.
Benchmarks and standing
Qwen3-Coder-480B-A35B-Instruct reports the following benchmark positions at release:
- SWE-Bench Pro: 38.7 percent
- Terminalbench 2.0: 23.9 percent
- Evasion Bench: 78.16 percent
The published model card cites comparison against Claude Sonnet on the agentic-coding axes, framing the model as competitive with the closed-source frontier on the kind of tasks IDE-integrated agents handle. Independent benchmark reproductions and the broader competitive-coding leaderboard positions are referenced in the technical report rather than itemised on the model card; the headline numbers above are the principal first-party claims.
The competitive picture at release in May 2025 was that Qwen 3 Coder sat in the leading group of open-weights coding specialists alongside DeepSeek Coder. By May 2026, the broader category has moved on: newer reasoning-and-coding models including GLM-5.1 (SWE-Bench Pro 58.4 percent), Kimi K2.5 (SWE-Bench Verified 76.8 percent, SWE-Bench Pro 50.7 percent), and Qwen 3.6 (SWE-bench Verified 73.4 percent) post materially higher figures on overlapping evaluations. Qwen 3 Coder retains relevance as the specialized coding entry with the longest established deployment history and the deepest integration into the Qwen Code and CLINE tooling ecosystems.
Benchmark leadership in the open-weights coding category turns over on roughly a quarterly cadence. The successor in the Qwen 3 family has been the Qwen 3.6 line, which moves coding capability into the general-purpose multimodal flagship rather than maintaining a dedicated Coder branch.
Access and pricing
Qwen 3 Coder 480B-A35B Instruct ships under Apache 2.0, permitting research and commercial use without per-token licensing. Distribution channels:
- Hugging Face Hub as the primary release repository.
- Qwen Code CLI as the first-party command-line coding agent.
- CLINE as the principal third-party IDE agent integration.
- Fireworks AI hosts the model in the LLM catalog.
- Novita and other inference providers offer paid API access.
- Local deployment frameworks: Ollama, LMStudio, MLX-LM (Apple Silicon), llama.cpp, KTransformers, vLLM, SGLang, Docker Model Runner. Minimum transformers version 4.51.0 for full feature support.
- HuggingChat offers the hosted chat interface.
Recommended sampling parameters from the model card: temperature 0.7, top-p 0.8, top-k 20, max tokens 65,536. The model operates in non-thinking instruct mode only.
Comparison
- Qwen 3.6 (Alibaba Qwen). The successor general-purpose generation. Qwen 3.6's flagship 35B-A3B variant reports stronger SWE-bench Verified (73.4 percent) than Qwen 3 Coder's SWE-Bench Pro (38.7 percent) at a much smaller active-parameter count, reflecting the rapid pace of coding-benchmark improvement in the intervening year.
- GLM-5.1 (Z.ai). The principal newer open-weights coding specialist. GLM-5.1's SWE-Bench Pro of 58.4 percent materially exceeds Qwen 3 Coder's 38.7 percent on the same benchmark, reflecting the year's progress.
- Kimi K2.5 (Moonshot AI). The multimodal frontier-tier peer with strong coding capability (SWE-Bench Verified 76.8 percent, SWE-Bench Pro 50.7 percent).
- DeepSeek V4 (DeepSeek). The general-purpose Chinese frontier-tier peer. DeepSeek Coder is the dedicated coding entry in the family, sitting in the same competitive frame as Qwen 3 Coder at the time of release.
- Claude Sonnet (Anthropic). The closed-source coding-frontier peer that the Qwen 3 Coder model card explicitly compares against on agentic-coding tasks.
The competitive question for Qwen 3 Coder in mid-2026 is whether the dedicated coding-line branding still produces meaningful product differentiation, or whether the move to coding capability in general-purpose flagships (Qwen 3.6, GLM-5.1, Kimi K2.5) has consolidated the category around the broader model lines.
Outlook
Open questions for the next 6 to 18 months:
- Successor in the dedicated Coder line. Whether Alibaba Qwen releases a Qwen 3.6 Coder or a Qwen 4 Coder variant, or whether the general-purpose Qwen 3.6 family's coding capability eliminates the need for a dedicated Coder branch, is the central roadmap question.
- Independent SWE-Bench Pro reproduction at scale. The 38.7 percent SWE-Bench Pro figure is first-party. Comparison against the same benchmark for the newer peer cohort (GLM-5.1, Kimi K2.5, Qwen 3.6) would clarify the current relative position.
- Qwen Code and CLINE adoption metrics. The application-layer surfaces (Qwen Code CLI, CLINE IDE integration) are the principal product manifestations of the Qwen 3 Coder model. Public adoption metrics or feature-comparison data against the closed-source coding agents (Claude Code, Cursor Composer, GitHub Copilot) would clarify the competitive position.
- Apache 2.0 license stability. The Apache 2.0 release is more permissive than the modified MIT licenses applying to MiniMax and Kimi K2.5. Whether subsequent Qwen Coder releases maintain Apache 2.0 or move to more restrictive licensing is a structural question for the enterprise-adoption pattern.
- Tool-call format standardisation. The custom function-call format limits the model's drop-in compatibility with broader agent frameworks. Whether Alibaba Qwen converges on the OpenAI tool-call schema in a future release, or maintains the custom format as a deliberate ecosystem-anchoring choice, is unannounced.
Sources
- Hugging Face: Qwen3-Coder-480B-A35B-Instruct. Primary model card with architecture, benchmark, and distribution details.
- Qwen 3 Coder technical paper. The detailed architecture and training description.
- Fireworks AI: model catalog. Hosted-inference availability.
- Companion profile: Alibaba Qwen for the broader Qwen family roadmap.
- Companion model: Qwen 3.6 for the successor general-purpose generation with strong coding capability.
- Companion model: GLM-5.1 for the principal newer open-weights coding specialist.
- Companion model: Kimi K2.5 for the multimodal frontier-tier peer with strong coding capability.
