DEV.co
Open-Source LLM · Qwen

Qwen3-235B-A22B-Thinking-2507-FP8

Qwen3-235B-A22B-Thinking-2507-FP8 is an Apache-2.0-licensed, open-source 235B-parameter mixture-of-experts language model from Alibaba's Qwen team. It is FP8-quantized for deployment efficiency and emphasizes reasoning and extended thinking for complex problem-solving. The model supports a native 256K context length and can generate up to 81K tokens for challenging reasoning tasks. It is available without gating and has been benchmarked against leading closed-source models (GPT-4O, Claude, Gemini, DeepSeek-R1) on reasoning, coding, and knowledge tasks.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507-FP8
235.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
44.4k
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
Downloads44.4k
Likes86
Last updated2025-07-30
SourceQwen/Qwen3-235B-A22B-Thinking-2507-FP8

What Qwen3-235B-A22B-Thinking-2507-FP8 is

Qwen3-235B-A22B-Thinking-2507-FP8 is a causal language model with Mixture-of-Experts (MoE) architecture: 235B total parameters with 22B activated at inference; 94 layers; 128 experts (8 activated per forward pass); GQA with 64 Q-heads and 4 KV-heads. Quantized to FP8 (block size 128) for memory and inference efficiency. Context window: 262,144 tokens (native). Requires transformers ≥4.51.0 for proper tokenizer support. Compatible with vLLM (≥0.8.5), SGLang (≥0.4.6.post1), and local tools (Ollama, LMStudio, llama.cpp, KTransformers). Deployment frameworks must support the 'qwen3_moe' architecture. Model enforces internal reasoning (thinking) via default chat template; output includes explicit </think> tokens.

Quickstart

Run Qwen3-235B-A22B-Thinking-2507-FP8 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-FP8")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-grade problem-solving in STEM domains

Benchmarked on AIME25 (92.3), HMMT25 (83.9), and GPQA (81.1). Suitable for mathematical theorem proving, scientific analysis, and multi-step academic problem-solving where detailed reasoning justification is valuable.

Production code generation and software engineering

LiveCodeBench v6 score of 74.1% and CFEval (2134) indicate strength in real-world coding tasks. Useful for code-generation pipelines and AI-assisted development when reasoning transparency is beneficial.

Self-hosted agentic and tool-calling applications

Open-source license and gating-free distribution enable deployment in enterprise environments. Qwen-Agent integration supports tool definition via MCP config; suitable for custom LLM apps, RAG systems, and private reasoning agents.

Running & fine-tuning it

ESTIMATE (unverified; requires independent testing): FP8 quantization (~120–150 GB VRAM for full model on single GPU; tensor-parallel sharding recommended for 4+ GPUs). Model card recommends context length ≥131K for reasoning; OOM mitigation may require reducing context or using vLLM/SGLang with paged attention. Inference framework (vLLM/SGLang) essential for memory-efficient serving. Exact VRAM and throughput depend on quantization, context length, and inference framework tuning.

Not explicitly documented in the provided model card. Fine-tuning feasibility (LoRA/QLoRA) is Unknown. The model is positioned as a pre-trained and post-trained base model; no official fine-tuning guidelines, datasets, or target use cases are stated. Organizations considering fine-tuning should consult the GitHub repository (https://github.com/QwenLM/Qwen3) or Qwen documentation for supported methodologies.

When to avoid it — and what to weigh

  • Latency-critical real-time applications — Model size (235B, 22B active) and generated output up to 81K tokens for complex reasoning create high latency. Not suitable for sub-second response requirements without sophisticated serving optimization.
  • Single-token or streaming constrained environments — Thinking mode is mandatory and cannot be disabled. Output always includes internal reasoning tokens (</think>), increasing response length. Environments requiring fixed-size responses or minimal token generation overhead should evaluate alternatives.
  • Limited VRAM or edge/mobile deployments — 235B parameters require substantial GPU memory even in FP8 (model card recommends context length ≥131K for reasoning; OOM issues reported at reduced context). FP8 quantization helps but full model unsuitable for edge devices or small inference clusters.
  • Non-English-only use cases with minimal multilingual testing — While MultiIF and multilingual benchmarks are reported, testing outside English is less comprehensive than English. Organizations prioritizing non-English performance should validate on actual workloads.

License & commercial use

Apache License 2.0 (OSI-compliant permissive open-source license). Allows commercial use, modification, and distribution with attribution and no warranty. No artificial restrictions; model is ungated and downloadable without approval.

Apache-2.0 is an OSI-approved permissive license permitting commercial use, modification, and sublicensing. No additional restrictions stated in the model card. Commercial deployment is permitted under Apache-2.0 terms (attribution required; no liability accepted). Verify license compliance with legal counsel for production deployments. No proprietary usage restrictions or commercial licensing terms are implied.

DEV.co evaluation signals

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

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

Typical LLM security considerations apply: (1) Model outputs should not be treated as ground truth without validation; (2) Long-context processing (256K tokens) and extended reasoning could enable prompt injection or exfiltration attacks if inputs are untrusted; (3) FP8 quantization does not introduce documented security risks but reduces precision and may alter adversarial robustness vs. BF16 variant; (4) Open-source model weights are publicly available, enabling offline analysis but also adversarial reverse-engineering; (5) Thinking-mode output includes internal reasoning, which may leak model behavior or confidential problem structure if logs are exposed. Recommended: validate inputs, isolate reasoning traces, and audit logs in production deployments.

Alternatives to consider

DeepSeek-R1 (open-source, similar thinking-based reasoning)

DeepSeek-R1 also emphasizes extended reasoning and is open-source, with comparable benchmarks on reasoning tasks. Smaller model sizes available (7B–671B range) may be more memory-efficient for some use cases. Requires independent evaluation of commercial-use license and community support.

OpenAI GPT-4O / GPT-4-Turbo (closed-source, highest reasoning benchmarks)

Proprietary models with higher AIME25 (92.7), HMMT25 (82.5), and broader real-world deployment support. No self-hosting or fine-tuning; API-based pricing. Preferred for enterprises avoiding open-source dependency risk but accept vendor lock-in and per-token costs.

Llama 3.1-405B (open-source, larger unified model, no enforced thinking)

Software development agency

Ship Qwen3-235B-A22B-Thinking-2507-FP8 with senior software developers

Qwen3-235B-A22B-Thinking-2507-FP8 is production-ready for enterprises building AI agents, research tools, and code-generation systems. Start with vLLM or SGLang for efficient inference, or consult our team to architect your reasoning LLM pipeline.

Talk to DEV.co

Related open-source tools

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

Qwen3-235B-A22B-Thinking-2507-FP8 FAQ

Can I use this model commercially without additional licensing?
Yes, under Apache-2.0 terms. The license permits commercial use, modification, and distribution with attribution and no liability warranty. No additional commercial license purchase is required. Verify compliance with your legal team for production deployments.
What GPU memory do I need to run Qwen3-235B-A22B-Thinking-2507-FP8?
Exact VRAM is not specified in the model card. FP8 quantization reduces size vs. BF16, but 235B parameters with 22B activation still require substantial GPU RAM. Model card recommends context length ≥131K for reasoning; OOM issues are common at reduced context. Use vLLM/SGLang with paged attention, tensor parallelism (4+ GPUs), or quantization further. Run benchmarks on your hardware before production deployment.
Is fine-tuning (LoRA, QLoRA) supported for this model?
Not explicitly documented in the model card. Fine-tuning feasibility and best practices are unknown. Consult the GitHub repository (https://github.com/QwenLM/Qwen3) or official Qwen documentation for supported fine-tuning methodologies and tooling.
How does the 'thinking' mode work, and can I disable it?
The model defaults to reasoning/thinking mode via its chat template. Output includes </think> tokens marking the end of internal reasoning. Thinking mode is mandatory; it cannot be disabled via a flag. Output length increases due to reasoning tokens (up to 81K for complex tasks). This is by design for transparent problem-solving but increases latency and token usage.

Work with a software development agency

DEV.co helps companies turn open-source tools like Qwen3-235B-A22B-Thinking-2507-FP8 into production software. Our software development services cover the full lifecycle — architecture, web development, integration, and maintenance — delivered by software developers and web developers who ship. Engage our software development agency to implement or customize it for your open-source llms stack.

Deploy Advanced Reasoning with Qwen3

Qwen3-235B-A22B-Thinking-2507-FP8 is production-ready for enterprises building AI agents, research tools, and code-generation systems. Start with vLLM or SGLang for efficient inference, or consult our team to architect your reasoning LLM pipeline.