DEV.co
Open-Source LLM · Qwen

Qwen3-235B-A22B-Thinking-2507

Qwen3-235B-A22B-Thinking-2507 is a 235B-parameter open-source LLM from Alibaba's Qwen team, featuring a mixture-of-experts (MoE) architecture with 22B active parameters. It is designed for complex reasoning tasks (mathematics, coding, science) with native 256K context support and an embedded thinking mechanism that generates internal reasoning chains before producing output. The model is freely available under Apache 2.0, but has high computational requirements and specialized deployment needs.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507
235.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
46.2k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperQwen
Parameters235.1B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads46.2k
Likes407
Last updated2025-08-17
SourceQwen/Qwen3-235B-A22B-Thinking-2507

What Qwen3-235B-A22B-Thinking-2507 is

Causal language model with 94 transformer layers, 128 MoE experts (8 active per token), grouped query attention (64 Q-heads, 4 KV-heads), and 262,144-token native context window. Trained on both pretraining and post-training stages. Inference requires thinking mode (outputs reasoning before response). Compatible with Hugging Face transformers (>=4.51.0), vLLM (>=0.8.5), SGLang (>=0.4.6.post1), Ollama, LMStudio, and llama.cpp. No embedding layer parameters counted separately (234B non-embedding).

Quickstart

Run Qwen3-235B-A22B-Thinking-2507 locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="Qwen/Qwen3-235B-A22B-Thinking-2507")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Research & Scientific Problem-Solving

Strong performance on GPQA (81.1), MMLU-Pro (84.4), and academic benchmarks. Suited for literature synthesis, hypothesis generation, and technical document analysis requiring deep reasoning.

Competitive Programming & Algorithm Development

Achieves 74.1% on LiveCodeBench and 2134 on CFEval. Effective for code generation, algorithm design, and multi-step coding tasks that benefit from explicit reasoning traces.

Complex Reasoning & Agentic Workflows

Designed for multi-turn reasoning and tool-use scenarios. Recommended for customer service agents, enterprise search, and task automation requiring interpretable decision-making (thinking output visible).

Running & fine-tuning it

ESTIMATE: ~470 GB GPU VRAM for full bfloat16 inference with 8× tensor parallelism (59 GB per GPU). Requires high-bandwidth multi-GPU interconnect (NVLink or equivalent). For 262K context, memory scales linearly; recommended minimum is 8× H100 80GB or A100 80GB. Context length >131K strongly advised; reducing below 131K may degrade reasoning quality. Quantized versions (int8, fp8) may reduce VRAM to ~235 GB, but throughput/accuracy trade-offs require testing.

Model card does not specify LoRA or QLoRA feasibility. Given 235B size and MoE architecture, full fine-tuning is prohibitively expensive. Unknown: LoRA applicability to MoE layers, instruction tuning data format compatibility, or supported quantization schemes for training. Recommend consulting Qwen GitHub and official documentation before attempting fine-tuning; consider prompt engineering or RAG as alternatives.

When to avoid it — and what to weigh

  • Low-latency or Real-time Applications — Model produces extended thinking output (up to 81,920 tokens on reasoning tasks), resulting in high latency. Unsuitable for interactive chatbots, voice assistants, or sub-second response requirements.
  • Resource-Constrained Environments — 235B parameters require multi-GPU deployment (model card recommends 8× tensor parallelism). Cannot run on single GPUs or edge devices without quantization; even quantized versions demand high VRAM. Requires >131K context length for optimal reasoning.
  • Non-English or Specialized Domain Tasks Without Retraining — While multilingual capabilities exist (80.6% MultiIF), domain-specific performance (legal, medical) is not documented. May require fine-tuning or RAG augmentation for specialized use cases.
  • Deterministic or Cost-Sensitive Inference — Thinking mechanism increases token generation cost significantly. Token-counting systems must account for hidden thinking tokens. Not suitable for applications requiring strict per-inference budgets.

License & commercial use

Apache 2.0 license. Permissive OSI-compliant open-source license allowing modification, redistribution, and commercial use with attribution. No proprietary restrictions or gating. Model weights are publicly available on Hugging Face without login requirement (gated: false).

Apache 2.0 explicitly permits commercial use. No restrictions on derivative works, proprietary applications, or SaaS deployment observed. However: (1) verify compliance with any underlying training data licenses (not stated in card), (2) MoE architecture and thinking mechanism may require specific deployment infrastructure licensing (e.g., GPU software licenses), and (3) production deployment should include model governance and output validation. No vendor lock-in risk; self-hosted deployment fully supported.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Standard LLM risks apply: (1) thinking output is visible to end users—may leak intermediate reasoning, proprietary logic, or sensitive data passed through prompts; (2) no adversarial robustness testing documented; (3) no explicit safeguards against prompt injection or jailbreaking mentioned; (4) tool-calling capability (agentic mode) expands attack surface if not properly sandboxed; (5) MoE routing behavior not audited for timing side-channels. Recommend: treat thinking output as sensitive, implement prompt filtering, sandbox tool execution, and regularly audit generated content. No security incidents or penetration test results disclosed.

Alternatives to consider

DeepSeek R1 (70B/671B open-source)

Comparable reasoning model with lower parameter count (70B option for resource-constrained settings). Slightly lower scores on some benchmarks but lower deployment cost. Closed reasoning format (non-visible thinking).

OpenAI o3 / o4-mini (API-only)

Higher absolute benchmark scores (e.g., 94.9 vs. 93.8 on MMLU-Redux, 88.9 vs. 92.3 on AIME25) but proprietary, requires API spend, and no local deployment. No control over model updates or data privacy.

Claude 3.5 Sonnet Thinking (API-only)

Anthropic's thinking model with strong alignment and tool use. No open-source weights; API-based only. Suitable if cost and vendor lock-in are acceptable trade-offs for enterprise support and safety.

Software development agency

Ship Qwen3-235B-A22B-Thinking-2507 with senior software developers

Qwen3-235B-A22B-Thinking-2507 is production-ready for R&D, scientific computing, and complex task automation. Verify your infrastructure supports multi-GPU deployment, then contact our AI platform team to architect a cost-effective self-hosted or cloud solution.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Qwen3-235B-A22B-Thinking-2507 FAQ

Can I use Qwen3-235B-A22B-Thinking-2507 commercially?
Yes. Apache 2.0 license permits commercial use, SaaS deployment, and proprietary applications. No royalties or licensing fees. Ensure compliance with training data licenses (not documented in card) and implement appropriate output monitoring for production use.
What GPU hardware do I need to run this model?
Minimum: 8× H100 80GB or A100 80GB GPUs for full bfloat16 inference (~470 GB total VRAM). Single-GPU deployment is not feasible without severe quantization trade-offs. vLLM or SGLang with 8× tensor parallelism is recommended for production.
Why is the model so slow? How do I reduce latency?
The model is optimized for reasoning quality, not speed. It generates internal thinking tokens (up to 81,920 on complex tasks) before producing output, which increases total latency. Reduce via: (1) shorter max_new_tokens, (2) simpler prompts (less thinking needed), (3) disabling thinking mode if reasoning is not required (not explicitly supported; check Qwen-Agent). Expect 10-30s+ for complex problems.
Does the model support fine-tuning?
Not documented in the card. Full fine-tuning is infeasible (235B params, high cost). LoRA/QLoRA compatibility with MoE layers unknown. Contact Qwen or consult the GitHub repo for guidance. Prompt engineering and RAG are recommended alternatives.

Software developers & web developers for hire

Adopting Qwen3-235B-A22B-Thinking-2507 is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy Enterprise Reasoning AI?

Qwen3-235B-A22B-Thinking-2507 is production-ready for R&D, scientific computing, and complex task automation. Verify your infrastructure supports multi-GPU deployment, then contact our AI platform team to architect a cost-effective self-hosted or cloud solution.