Qwen-Agent
Native Qwen agent framework with MCP, RAG, code interpreter, and Gradio UI.
Qwen3 235B A22B is the open-weight model that closes the gap with closed frontier labs. Released April 28, 2025 by the Qwen team at Alibaba, it ships under Apache 2.0 with no commercial caveats.
Specs that matter when picking the model and the hardware.
235B total, 22B active per token
MoE, 128 experts, 8 active per token
128K native, 262K on Instruct-2507
Apache 2.0
April 28, 2025 (Instruct-2507 in July 2025)
119 languages and dialects
Hybrid Thinking and Non-Thinking
143GB
112GB
15-22
120-180 batched
Three tiers, real SKUs, real throughput numbers.
Throughput: 15-22 tokens/sec. Best for: Single-user power workflows, code generation, deep reasoning on private docs.
Throughput: 40-70 tokens/sec. Best for: 10-30 user firms, vLLM batched serving, MoE expert offload.
Throughput: 120-300 tokens/sec. Best for: 100+ concurrent users, multilingual ops, agentic workflows.
GGUF, AWQ, FP8, INT4.
IQ4_XS for 128GB Macs (close to Q4_K_M quality, fits with 40K context), Q4_K_M for 256GB+ Macs, FP8 for H100 and H200, AWQ INT4 for Blackwell consumer cards
MoE models tolerate quantization slightly better than dense models because expert routing dilutes per-weight error. IQ4_XS loses around 3 percent on MMLU-Pro versus FP16. FP8 is the production sweet spot on Hopper and Blackwell, near-lossless and 2x throughput versus FP16.
File size: 112GB.
File size: 143GB.
File size: 118GB.
File size: 235GB.
Native tool calling with the strongest reliability of any open model in 2026.
Native option from Alibaba: pip install -U qwen-agent[gui,rag,code_interpreter,mcp]. Encapsulates tool-calling templates and parsers internally. Pass model_server as the vLLM /v1 URL. Bundles MCP, code interpreter, RAG, and Chrome extension.
OpenAI Compatible provider. Tool calling on Qwen3 is the most stable in the open model class, rarely hallucinating calls or dropping parameters. Plan and Act mode both work. The Thinking-2507 variant adds reasoning steps for deeper plans, the Instruct-2507 is faster for routine work.
Aider Polyglot leaderboard puts Qwen3 235B A22B at 65.3 percent, near the top of self-hostable open models. Configure --openai-api-base to vLLM endpoint. Edit-format mode works cleanly across Python, TypeScript, Go, Rust, and Java.
Add to config.yaml with provider: openai and apiBase to your vLLM. Declare capabilities: [tool_use]. Agent mode auto-detects native tool calling. The 256K context on the Thinking variant unlocks Continue's full repo-context mode.
Strong fit. A 235B RFT checkpoint of the Instruct-2507 base hit 61.7 percent Pass@1 (74.3 percent Pass@5 at 500 turns) on SWE-bench Verified using OpenHands trajectories per Nebius, beating the 30B coding specialist (50.3 percent) at half the parameters of Qwen3-Coder-480B. Set LLM_BASE_URL to vLLM. Best autonomous-loop performance of any model in this batch.
Cline fork. Multi-mode (Code, Architect, Ask, Debug). The Architect mode + Thinking-2507 variant is a strong combination because Qwen3's reasoning trace surfaces architectural tradeoffs cleanly. Same OpenAI-compatible config.
Claude Code accepts OPENAI_BASE_URL override per the unofficial guides. Qwen3 235B A22B is the open model that comes closest to Claude Sonnet on Claude Code's own SWE-bench eval. Watch tool-format compatibility, some Claude Code internal tools assume Anthropic-flavored function calls.
Where this model + agent stack ships real work.
256K context (Thinking-2507 variant) makes Qwen3 235B A22B one of the strongest open models for deep research.
Built-in RAG, code interpreter, MCP. Single pip install gives a working agent stack. Best research starting point if the buyer is committed to the Qwen line.
Point OPENAI_BASE_URL at vLLM. Qwen3's instruction following (Qwen 3.5 hits IFEval 92.6) means GPT Researcher's planning prompts are followed precisely. Reports run 3 to 5 minutes for 5 to 10 sources.
Generic OpenAI provider works. Workspace RAG plus @agent mode covers 80 percent of buyer use cases. Qwen3's tool reliability means @agent rarely fails on multi-tool sequences.
Default chat UI. Qwen3's hybrid thinking toggle exposes nicely via Open WebUI's per-conversation settings (chat_template_kwargs.enable_thinking). Buyers can flip reasoning on for hard questions, off for fast chat.
Production research orchestration. Qwen3's reliable tool calling is the differentiator: LangGraph state transitions trigger correctly without retry logic. Pair with LangSmith tracing.
Role-based research crews. Qwen3 235B A22B's strong instruction following means CrewAI's role and goal text is honored without prompt-engineering overhead.
Best retrieval pairing. Use LlamaIndex query engines as tools inside a Qwen-Agent or LangGraph workflow. The 256K context lets you keep a much larger retrieved chunk window than typical RAG.
Prompt optimization. DSPy's compile step is unusually effective on Qwen3 because the model follows structured few-shot patterns precisely. Use DSPy to optimize the question-decomposition prompt, then deploy via LangGraph.
Where this model + research agent stack delivers reports your team actually uses.
Pick the framework by workflow shape.
Native Qwen agent framework with MCP, RAG, code interpreter, and Gradio UI.
Production stateful orchestration. Best paired with Qwen3's reliable tool calling.
Role-based crews. Fastest to demo with Qwen3's instruction following.
GroupChat conversations. Useful for Qwen3 hybrid mode debate patterns.
Prompt compilation. Strong with Qwen3's structured-format adherence.
Type-safe agents. Qwen3's structured output reliability is a clean fit.
Retrieval layer. Pair with the 256K Qwen3 context for huge retrieved chunks.
Three buyer profiles where this model is the right answer.
Task: Cross-border contract review in English, French, German, Mandarin, plus discovery. Outcome: 119-language coverage replaces three separate translation tools. Apache 2.0 means redistribution to client portals is clean.
Task: Code generation, code review, agentic refactoring across multiple languages and frameworks. Outcome: 70.7 LiveCodeBench v5 and 2,056 CodeForces beats every dense open model. Hybrid Thinking mode for hard bugs, Non-Thinking for autocomplete.
Task: Literature synthesis, hypothesis generation, GPQA-class STEM Q&A on internal documents. Outcome: SuperGPQA leadership over DeepSeek-V3 and Llama-4 on expert STEM. 256K context handles full clinical trial dossiers.
Feedback from the people who were in the room when decisions were made.
I used to spend every Sunday reconciling contractor invoices. Now they match automatically. I got my weekends back.
Three drivers called out on the same morning. The system reassigned every route before I finished my coffee. That used to take two hours of phone calls.
They did not ask us to trust a rewrite plan. They proved one critical workflow first, then scaled from evidence.
Where this model wins, where it loses.
DeepSeek V3.2 (671B MoE, 37B active) wins on math and code at the absolute top end. Qwen3 235B wins on memory footprint (143GB Q4 versus 376GB Q4) and on hardware accessibility for any firm without an 8x H100 budget.
Qwen3 235B wins on reasoning, coding, multilingual, and benchmarks across the board. Llama 3.3 70B wins on serving simplicity, hardware floor, and ecosystem maturity. Pick Llama for 8 to 15 person firms, Qwen3 235B for 30+ where quality is the constraint.
Mistral Large 3 wins on Apache 2.0 plus first-class multimodal vision. Qwen3 235B wins on hardware footprint (143GB versus 370GB at Q4) and on the size-of-firm that can self-host without an enterprise GPU node.
Day 4 vLLM (or SGLang) serves the model with tool calling and reasoning parser.
vLLM serves Qwen3-235B-A22B-Instruct-2507 with tensor-parallel-size 8, max-model-len 131072 (or 256K on Thinking-2507), --enable-reasoning --reasoning-parser deepseek_r1 for the Thinking variant. FP8 variant fits on 8xH100.
Editor agents (Aider, Cline, Continue.dev, Roo Code) wired to /v1. OpenAI-compatible config across all. Tool calling is the most stable of any open model so almost no retry logic needed.
Open WebUI plus AnythingLLM live. Hybrid thinking toggle exposed per conversation via chat_template_kwargs.enable_thinking. Workspaces isolate sensitive doc sets.
Qwen-Agent installed (pip install -U qwen-agent[gui,rag,code_interpreter,mcp]). MCP servers wired for filesystem, time, fetch, and your internal APIs.
LangGraph workflows for production research. Three workflows shipped: deep-research, code-review, customer-intake. LangSmith tracing on.
OpenHands repo-autonomy demoed against three real tasks. Expected SWE-bench-class pass rate around 60 percent given the 235B Instruct base.
Evidence pack with tool-call success rates, throughput numbers, accuracy spot-checks, runbook, and agent inventory delivered.
Fixed fee.
Confirm hardware tier. 128GB Mac is borderline (IQ4_XS only). 256GB+ Mac or multi-GPU is the real floor. Bench Thinking versus Non-Thinking on actual workload.
Serve via vLLM with FP8 (GPU) or MLX with Q4 (Mac). Wire OpenAI-compatible API. Configure thinking-mode toggle by route or by user role.
RAG with multilingual BGE-M3 embeddings, Qdrant. SSO, audit logging, RBAC, expert-routing telemetry for cost monitoring.
Pen-test cleared. Apache 2.0 NOTICE filed. Cutover. No attribution string required in product UI.
235B total, 22B active per token (MoE) Parameters
Fixed fee.
Day 1 you see the exact model, the exact hardware, the exact agent stack, the exact cost math, and the exact compliance posture.
Yes. Apache 2.0, no MAU cap, no attribution-in-product requirement, no paid tier. You can fine-tune, redistribute, ship in commercial products, and embed in client deliverables without contacting Alibaba. The only requirement is preserving the Apache 2.0 NOTICE file in source distributions, which is a one-line legal task.
Mac Studio M3 Ultra 256GB at Q4_K_M is the cheapest viable path, 15 to 22 tokens per second decode for a single user. 128GB Mac requires IQ4_XS and runs out of room past 40K context. For multi-user production, target 2x RTX Pro 6000 Blackwell 96GB or a 4x H200 node with FP8.
Thinking mode runs an internal chain-of-thought before answering, slower but stronger on math, code, and multi-step reasoning. Non-Thinking is direct response, faster, suited for chat, drafting, and tool calls. The 2507 Instruct variant is Non-Thinking by default. Route hard problems to Thinking and commodity tasks to Non-Thinking and your tokens-per-dollar improves 3 to 5x.
DeepSeek V3.2 wins on hardest reasoning and IMO-class math. Qwen3 235B wins on hardware floor (143GB versus 376GB at Q4), multilingual breadth, and serving simplicity. For any firm without an 8x H100 cluster sitting idle, Qwen3 235B is the practical pick.
Qwen3's post-training pipeline included RL across 20+ general-domain tasks specifically to strengthen agent capabilities, instruction following, and format following. You get the lowest tool-call hallucination rate of any open model in the 2026 benchmarks, and the chat template natively encodes the OpenAI-compatible tool format plus MCP. In Pineido sprints we see noticeably fewer retry-loop bugs on Qwen3 versus Llama or Mistral.
Both. Pineido deploys the Instruct-2507 as the default endpoint because it is faster and skips the thinking tokens. We deploy Thinking-2507 on a second port for hard problems. The two models share the architecture so vLLM serves them side by side. Cline, Roo Code, and Open WebUI can route per task. The Thinking variant scores ~7 to 12 points higher on hard reasoning benchmarks but is roughly 3x slower per response.
ByKostadin StambolievFounder and Imaginator
Pineido, based in Strumica, North Macedonia, is a software company that builds websites, online shops, web and mobile apps, business automation and AI for companies at home and abroad. Work ships in 30-day sprints at a fixed price, and the client owns the code.
Published Updated