DEV.co
Open-Source LLM · Qwen

Qwen3-235B-A22B

Qwen3-235B-A22B is a 235-billion-parameter mixture-of-experts (MoE) language model from Alibaba's Qwen team. It activates only 22B parameters per inference step, balancing capability with efficiency. The model uniquely supports toggling between 'thinking mode' (for complex reasoning) and 'non-thinking mode' (for fast dialogue) within a single model, and claims strong performance on math, coding, reasoning, and multilingual tasks.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-235B-A22B
235.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
918.8k
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
Downloads918.8k
Likes1.1k
Last updated2025-07-26
SourceQwen/Qwen3-235B-A22B

What Qwen3-235B-A22B is

MoE architecture with 128 experts (8 active per token), 94 layers, GQA attention (64 Q heads, 4 KV heads), native 32K context (extendable to 131K via YaRN). Causal language model trained with post-training optimization. Requires transformers >= 4.51.0. Supports thinking/non-thinking mode switching via `enable_thinking` flag and soft `/think`/`no_think` user directives. Compatible with vLLM, SGLang, Ollama, llama.cpp, and KTransformers.

Quickstart

Run Qwen3-235B-A22B 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")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

Complex problem-solving with reasoning transparency

Use thinking mode for math competitions, logic puzzles, or coding tasks where intermediate reasoning steps improve solution quality. The model outputs chain-of-thought in `<think>` blocks for interpretability.

Fast, efficient conversational AI

Deploy with `enable_thinking=False` for customer support, FAQs, or real-time chat where low latency matters more than deep reasoning. Reduces computational overhead while maintaining Qwen2.5-level instruction quality.

Multilingual and cross-lingual applications

Leverage 100+ language support for translation, multilingual customer support, or instruction-following in non-English contexts. Model card states strong multilingual instruction-following capability.

Running & fine-tuning it

Estimated minimum 50–80GB VRAM for bfloat16 (no quantization) on multi-GPU setups (e.g., 8× H100/A100). Single-GPU inference requires quantization (AWQ, GPTQ, int8) to reduce footprint to ~30–50GB. Exact requirements depend on batch size and context length; consult deployment docs (vLLM, SGLang, TGI) for precision-specific guidance. MoE sparsity (22B active) reduces per-token compute vs. dense 235B models.

Card does not document LoRA, QLoRA, or fine-tuning procedures. Implications: (1) full fine-tuning likely requires multi-GPU distributed training; (2) parameter-efficient tuning (LoRA) feasibility is unknown. Recommend consulting Qwen GitHub/docs or community reports before committing to fine-tuning workflows. MoE architecture may complicate adapter placement.

When to avoid it — and what to weigh

  • Strict latency SLAs without reasoning disabled — Thinking mode incurs significant overhead (generates reasoning tokens before response). Non-thinking mode mitigates this, but avoid if you need sub-100ms response times across all queries without model-level switching.
  • Limited GPU memory (< 50GB VRAM) — Even with MoE sparsity, 235B parameters require substantial memory. Quantization (e.g., AWQ, GPTQ) or 8-bit inference is likely necessary. Single-GPU deployment is not feasible without aggressive quantization.
  • Requirement for deterministic, greedy outputs — Model card explicitly warns: 'DO NOT use greedy decoding' in thinking mode; use temperature=0.6, top_p=0.95 instead. If determinism is critical, evaluate carefully or disable thinking mode (requires trade-off).
  • Specialized domain fine-tuning with limited compute — No LoRA/QLoRA guidance in card. Full fine-tuning a 235B model requires multi-GPU clusters. Feasibility for domain adaptation on consumer hardware is unclear.

License & commercial use

Apache 2.0 license. Permissive OSI-compliant license allowing use, modification, and redistribution (including commercial use) under standard Apache 2.0 terms.

Apache 2.0 is a permissive, OSI-approved open-source license that explicitly permits commercial use. No gating, no restricted use restrictions in the license itself. However: (1) Verify downstream dependencies (e.g., transformers, vLLM) have compatible licenses; (2) review Qwen's terms of service on HuggingFace if using their inference endpoints; (3) ensure compliance with any third-party tokenizer or inference frameworks used in production. No proprietary model weights restrictions detected.

DEV.co evaluation signals

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

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

Standard LLM security considerations apply: (1) Model outputs may hallucinate or reflect training-data biases; (2) thinking mode generates internal reasoning tokens—validate that sensitive intermediate logic is not exposed unintentionally; (3) no explicit security audit or adversarial robustness data provided; (4) when deployed via APIs (vLLM, SGLang), enforce standard API security (rate limiting, input validation, access control); (5) quantized or pruned variants may alter adversarial properties (requires empirical validation).

Alternatives to consider

Llama 3.1 405B

Dense 405B model with strong reasoning and multilingual support. Larger parameter count but no MoE sparsity. Higher inference cost, stronger benchmarks reported by Meta, but less efficient per-token compute.

DeepSeek-R1

Reasoning-focused model with chain-of-thought capabilities. Smaller dense variant available (e.g., 70B). Comparable reasoning features, different training approach, less established multilingual support.

Qwen2.5-72B or Qwen2.5-32B

Smaller dense alternatives from same Qwen family. Lower VRAM requirements, simpler deployment, but no thinking/non-thinking mode toggle and reduced reasoning capability. Faster inference without reasoning overhead.

Software development agency

Ship Qwen3-235B-A22B with senior software developers

Start with vLLM or SGLang for optimized serving. Estimate GPU requirements based on your batch size and thinking-mode usage. Review the Qwen GitHub and deployment docs for production best practices.

Talk to DEV.co

Related open-source tools

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

Qwen3-235B-A22B FAQ

Can I use Qwen3-235B-A22B commercially?
Yes. Apache 2.0 license permits commercial use. Ensure downstream dependencies (transformers, vLLM, SGLang, etc.) are also commercially compatible, and review Qwen/Alibaba's terms of service if using their hosted endpoints.
What GPU setup do I need for production inference?
Minimum 50–80GB VRAM for multi-GPU (e.g., 2–8× H100/A100 for bfloat16). Single-GPU production requires quantization (AWQ/GPTQ) to fit ~30–50GB. Use vLLM or SGLang for optimized serving; exact specs depend on batch size and thinking-mode usage.
How do I disable thinking mode for faster responses?
Set `enable_thinking=False` in `tokenizer.apply_chat_template()` or add `/no_think` to user prompts. Non-thinking mode behaves like Qwen2.5-Instruct with faster inference. Use temperature=0.7, top_p=0.8 for non-thinking mode.
Can I fine-tune this model for my domain?
Card does not document LoRA, QLoRA, or fine-tuning. Full fine-tuning requires multi-GPU distributed training and is resource-intensive. Check Qwen GitHub for community reports or fine-tuning guides before committing. MoE architecture may complicate parameter-efficient adaptation.

Custom software development services

Need help beyond evaluating Qwen3-235B-A22B? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy Qwen3-235B-A22B?

Start with vLLM or SGLang for optimized serving. Estimate GPU requirements based on your batch size and thinking-mode usage. Review the Qwen GitHub and deployment docs for production best practices.