Updated July 2026
Open Weights · Apache 2.0
Model Review
Alibaba's 480B-parameter coding specialist scores within 6 points of Claude Opus 5 on SWE-Bench — at roughly 1/40th the price. Here's the honest breakdown, benchmark numbers, and exactly when to reach for it instead.
1820Words
0 minRead Time
11Sections
Jul 2026Updated
$0.21 Claude Opus 5 /call → $0.003 Qwen3-Coder (now Qwen3-Coder-Next) /call
Affiliate Disclosure: OpenRouter is an affiliate partner — if you sign up through our link, we may earn a 20% recurring commission at no extra cost to you. All benchmarks, pricing data, and opinions are our own. We tested Qwen3-Coder (now Qwen3-Coder-Next) in production before writing this review.
This review is part of AIGetFree's AI coding tools coverage. For the terminal-native alternative, see our Claude Code Review. For the GUI editor, read the Cursor AI Review. For the complete developer stack, check the Ultimate AI Developer Stack.
The short version: Qwen3-Coder (now Qwen3-Coder-Next) is the open-source model that rivals GPT-5.6 for coding (70.6% SWE-bench, 3B active params) tasks. Free to self-host, cheap on OpenRouter, and shockingly capable for the price — with real gaps on multi-file refactors and long agent runs.
The first time you run Qwen3-Coder (now Qwen3-Coder-Next) on a real refactor — not a toy problem, not a HumanEval prompt, but a 400-line authentication module with three interdependent files and a failing test suite — something shifts. The model reads the stack trace, traces the dependency chain, and produces a diff that actually compiles. The whole exchange costs $0.003. The equivalent Claude Opus 5 call would have cost $0.21.
Open-source coding models have been promising "production-ready" for three years. Most delivered on benchmarks and failed in practice — verbose where you needed precision, confident where they should have been uncertain, brittle the moment a task required more than five tool-calling turns. Qwen3-Coder (now Qwen3-Coder-Next) is different. Not because it beats Claude Opus 5 — it doesn't, not consistently — but because the gap has narrowed to a point where the cost differential makes the choice obvious for most real workloads.
This review covers the architecture, the real benchmark numbers, production data from real pull requests, where the model breaks down, and the honest answer to the question every developer is actually asking: can I replace Claude with this and not notice? Sometimes yes. Sometimes no. Here's exactly when.
480BTotal Params
35BActive Per Token
300+Languages
$0.22Per 1M Input
Apache 2.0License
What Is Qwen3-Coder (now Qwen3-Coder-Next)?
Qwen3-Coder (now Qwen3-Coder-Next) is a coding-specialist large language model from Alibaba's Qwen team. It is a Mixture-of-Experts (MoE) model with 480 billion total parameters, of which only 35 billion activate per forward pass — 8 out of 160 experts fire for any given token. This architecture delivers frontier-class reasoning at a fraction of the inference cost of dense models.
The model was trained on 18 trillion code tokens across 300+ programming languages, with a native context window of 256K tokens and extrapolation support up to 1 million tokens. It is released under the Apache 2.0 license — meaning you can use it commercially, modify it, and self-host it without restriction.
🧠 Architecture
480B total / 35B active · MoE (8/160 experts) · 256K native context · 1M with extrapolation
📦 Access
OpenRouter ($0.22 / $1.80 per 1M) · Alibaba Cloud · Self-hosted (Hugging Face) · Ollama (quantized)
⚖️ License
Apache 2.0 — commercial use, modification, and self-hosting all permitted
🆚 vs Qwen3-Max
Qwen3-Max is the generalist. Qwen3-Coder (now Qwen3-Coder-Next) is the specialist — trained on code, agents, and tool-calling.
terminal — switching providers
# One-line switch from OpenAI to Qwen3-Coder (now Qwen3-Coder-Next) via OpenRouter
− base_url = "https://api.openai.com/v1"
+ base_url = "https://openrouter.ai/api/v1"
− model = "gpt-4.1"
+ model = "qwen/qwen3-coder"
# Everything else — LangChain, n8n, Cursor, Aider — works unchanged
Benchmark Results: The Numbers
All figures as of July 2026. SWE-Bench Verified uses standard agent scaffolds. Costs reflect OpenRouter pricing for Qwen3-Coder (now Qwen3-Coder-Next).
Output cost per 1M tokens
Qwen3-Coder (now Qwen3-Coder-Next)$1.60
Bars scaled to Claude Opus 5 = 100%. Qwen3-Coder (now Qwen3-Coder-Next)'s bar is widened slightly for visibility — the real ratio is closer to 47×.
| Metric |
Qwen3-Coder (now Qwen3-Coder-Next) |
GPT-5.4-Codex |
Claude Opus 5 |
| SWE-Bench Verified | 68.4% | ~70% | 74.6% 👑 |
| Aider Polyglot (pass@2) | 71.2% | — | 79.8% 👑 |
| HumanEval (pass@1) | 91.3% | 95.0% 👑 | 93.1% |
| LiveCodeBench | 64.7% | — | 70.2% 👑 |
| Input cost / 1M tokens | $0.40 🏆 | $2.50 | $15.00 |
| Output cost / 1M tokens | $1.60 🏆 | $15.00 | $75.00 |
| Context Window | 256K (1M ext.) | 128K | 1M 👑 |
| Open Weights | ✅ Apache 2.0 | ❌ Closed | ❌ Closed |
Real-World Performance: The Contra Collective Data
Benchmarks measure what models can do in controlled conditions. The Contra Collective PR Replay Corpus measures what they do on real pull requests — actual diffs, actual codebases, actual review criteria. The results are more nuanced than the leaderboards suggest.
| Metric |
Qwen3-Coder (now Qwen3-Coder-Next) |
Claude Opus 5 |
Gap |
| Technical Correctness | 3.8 / 5 | 4.4 / 5 | −0.6 |
| Code Style Match | 3.4 / 5 | 4.1 / 5 | −0.7 |
| Would-Merge Rate | 41% | 62% | −21pp |
| Cost Per PR | $0.018 🏆 | $1.27 | 70× cheaper |
| Rust Performance | +2.4pp vs Opus 🦀 | Baseline | Qwen wins |
The key insight: A 41% would-merge rate means Qwen3-Coder (now Qwen3-Coder-Next) produces code that needs review roughly 60% of the time. That's not a failure — that's a first draft. If your workflow treats LLM output as a starting point, the 70× cost advantage is decisive. If you need merge-ready code with minimal review, Claude Opus 5's 62% rate justifies the premium.
Where Qwen3-Coder (now Qwen3-Coder-Next) Shines
⚡
Short Tasks (<10 Turns)
Quality gap vs Claude closes to under 3pp on tasks under 10 tool-calling turns. For single-file edits and bug fixes, effectively peer-level at 47× lower output cost.
Speed💰
Cost-Sensitive Production
1,000 automated PRs/month: $18 with Qwen3-Coder (now Qwen3-Coder-Next) vs $1,270 with Claude. The math is not subtle. For coding agents at scale, this is the structural advantage of open-weight.
Value🦀
Rust Codebases
Outperforms Claude Opus 5 on Rust by 2.4pp — the only major language where it leads. If your stack is Rust-heavy, Qwen3-Coder (now Qwen3-Coder-Next) wins on both quality and cost.
Rust🔒
Self-Hosted / Air-Gapped
Apache 2.0 means zero data leaves your network. For financial services, healthcare, and defense — the only viable path to LLM-assisted coding at scale.
Privacy🤖
Agent Frameworks
Works natively with Cursor, Cline, Aider, Continue.dev, and OpenCode via OpenAI-compatible endpoint. Swap the model string — everything else runs unchanged.
Integrations🧪
Free Tier Prototyping
OpenRouter's free tier lets you test Qwen3-Coder (now Qwen3-Coder-Next) before spending a cent. Rate-limited but functional — validate whether it fits your workflow before committing.
Testing
Where It Falls Short
📁
Multi-File Refactors
18pp gap vs Claude on tasks touching 3+ files. Loses coherence across file boundaries — fixes file A but introduces inconsistencies in file B.
Weakness🔄
Long-Horizon Agentic Tasks
Degrades significantly on 20+ turn tasks. Begins losing context, repeating work, contradicting earlier outputs. Claude maintains coherence much longer.
Weakness🔷
Advanced TypeScript
Scored 1/10 vs Claude Sonnet's 8/10 on advanced type narrowing. Conditional types, infer keywords, and complex generics are a known weak point.
Weakness📝
Verbose Output
Ask for "diff only" and you'll often get a full file rewrite. Tends toward verbose output even when explicitly instructed otherwise.
Weakness🧪
Test Inference
Struggles to infer root cause from failing test output. Tends to over-engineer — rewriting entire functions when a two-line patch would suffice.
Weakness✍️
Non-Coding Tasks
A specialist, not a generalist. For writing, analysis, or reasoning outside code, use Qwen3-Max instead. Coding specialization costs general capability.
Limitation
Self-Hosting Reality
The Apache 2.0 license makes self-hosting legally straightforward. The hardware requirements make it practically complex. Here's the honest breakdown:
Full 480B (FP8)
480GB GPU memory
8× H100 or 4× H200
~$200K+ hardware
Quantized 30B (4-bit)
24GB VRAM
Single RTX 4090/5090
~$2,000 hardware
OpenRouter API
$0.22 / $1.80 per 1M
No hardware needed
Full 480B quality
terminal — running locally
# Ollama — quantized 30B on 24GB VRAM
ollama pull qwen3-coder:30b-q4_K_M
ollama run qwen3-coder:30b-q4_K_M
# vLLM — full 480B on 8×H100
vllm serve Qwen/Qwen3-Coder (now Qwen3-Coder-Next)-480B-A35B-Instruct \
--tensor-parallel-size 8 \
--enable-auto-tool-choice
vs. Other Open Models
| Model |
SWE-Bench |
Input $/1M |
Open |
Best For |
| Qwen3-Coder (now Qwen3-Coder-Next) 480B | 68.4% | $0.40 | ✅ | Coding agents, Rust, cost-sensitive |
| DeepSeek V3.2 | ~60% | $0.20 | ✅ | Ultra-cheap fallback |
| GLM-5.1 | ~70% | ~$0.40 | ✅ | Slightly better SWE-Bench |
| Qwen3-Max (general) | ~62% | ~$0.40 | ✅ | Writing, reasoning, non-code |
Honest Pros & Cons
✓ Pros
✓ 30–50× cheaper than Claude Opus 5 — structural cost advantage
✓ Apache 2.0 — self-host commercially, modify freely
✓ Beats Claude Opus 5 on Rust by 2.4pp
✓ OpenAI-compatible — works with Cursor, Cline, Aider, n8n
✓ 256K native context, 1M with extrapolation
✓ Quantized 30B runs on single 24GB GPU
✓ 68.4% SWE-Bench — within 6pp of best closed model
✗ Cons
✗ 41% would-merge rate — requires human review on most outputs
✗ Multi-file refactors: 18pp gap vs Claude
✗ Degrades on 20+ turn agentic tasks
✗ Advanced TypeScript: 1/10 vs Claude Sonnet's 8/10
✗ Verbose — ignores "diff only" instructions
✗ Full 480B self-hosting needs 8×H100
✗ No thinking mode — no visible reasoning steps
Frequently Asked Questions
Is Qwen3-Coder (now Qwen3-Coder-Next) better than GPT-5.6 for coding (70.6% SWE-bench, 3B active params)?
Depends on the task. On SWE-Bench they're within 2pp. GPT-5.4-Codex leads HumanEval (95% vs 91.3%). But on price-adjusted quality, Qwen3-Coder (now Qwen3-Coder-Next) wins by 5–6× — $0.40/$1.60 per million tokens vs $2.50/$15.00. For most production workloads, the quality difference is negligible and the cost difference is decisive.
Can I run Qwen3-Coder (now Qwen3-Coder-Next) locally?
Yes — two paths. Quantized 30B at 4-bit runs on a single 24GB GPU (RTX 4090/5090) via Ollama. Full 480B at FP8 needs 480GB GPU memory (8×H100). For most developers, the quantized 30B is the right local option, and OpenRouter is the right cloud option.
Does it work with Cursor, Cline, or Aider?
Yes — any tool with OpenAI-compatible API support works. Point Cursor to https://openrouter.ai/api/v1 with model qwen/qwen3-coder. For Aider use --openai-api-base. Nothing else changes.
Is it safe for production use?
Yes, with human review. The 41% would-merge rate means 6 in 10 outputs need revision before merging. That's a first-draft tool, not an autonomous committer. Treat Qwen3-Coder (now Qwen3-Coder-Next) output like a junior developer's PR — review required, but the draft is free.
What languages does it support?
300+ languages. Strongest: Python, TypeScript, Go, and Rust — where it actually beats Claude Opus 5. Weakest: C++ template metaprogramming and advanced TypeScript type narrowing. For mainstream languages, performance is competitive with frontier closed models.
Final Verdict
⭐⭐⭐⭐½
Qwen3-Coder (now Qwen3-Coder-Next) is not the best coding model — Claude Opus 5 holds that crown. But at 30–50× cheaper with open weights and Apache 2.0 licensing, it is the pragmatic choice for teams running coding agents at scale, for self-hosted deployments, and for anyone who treats LLM output as a first draft. The 6–9pp quality gap is real. The cost gap is structural and compounds at scale.
Best Value30–50× cheaper than Claude at near-comparable quality
Best for RustOutperforms Claude Opus 5 by 2.4pp on Rust tasks
Best for PrivacyApache 2.0 — self-host with full data ownership
Skip IfYou need autonomous multi-file refactors or 20+ turn agents
Try Qwen3-Coder (now Qwen3-Coder-Next) on OpenRouter →
Related Posts
You download the model, fire up Ollama, and watch it generate your first function.
You pull the Qwen3-Coder model and watch the download progress bar inch forward.
