DEV.co
Open-Source LLM · Qwen

Qwen3-235B-A22B-FP8

Qwen3-235B-A22B-FP8 is a 235-billion-parameter mixture-of-experts (MoE) language model from Alibaba's Qwen team, with only 22B parameters active per inference pass. It is quantized to FP8 for reduced memory footprint while maintaining quality. The model uniquely supports switchable thinking/reasoning mode (for complex tasks) and non-thinking mode (for fast dialogue). It supports 100+ languages and is optimized for instruction-following, agent-based tasks, and multilingual translation. Distributed under Apache 2.0 license with no access gating.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-235B-A22B-FP8
235.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
131.8k
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
Downloads131.8k
Likes94
Last updated2025-07-26
SourceQwen/Qwen3-235B-A22B-FP8

What Qwen3-235B-A22B-FP8 is

MoE architecture with 128 total experts, 8 activated per token. 94 transformer layers, GQA attention (64 Q-heads, 4 KV-heads). Native context window 32,768 tokens; expandable to 131,072 with YaRN. FP8 quantization uses fine-grained block-size-128 scheme. Requires transformers>=4.51.0. Inference supported via transformers, SGLang (>=0.4.6.post1), vLLM (>=0.8.5). Token-level thinking/non-thinking switching via chat template flag or soft-switch prompts (/think, /no_think).

Quickstart

Run Qwen3-235B-A22B-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-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

Complex Reasoning & Code Generation

Enable thinking mode for mathematics, algorithm design, and debugging. Model claims reasoning surpassing QwQ-32B and Qwen2.5-Instruct on these tasks.

Multilingual Dialogue & Translation

Leverage native 100+ language support and strong multilingual instruction-following for customer support, chatbots, and cross-lingual document processing.

Agent Systems & Tool Integration

Deploy with SGLang or vLLM for precise tool-use integration. Supports both thinking and non-thinking modes for agentic workflows, including complex planning tasks.

Running & fine-tuning it

ESTIMATE: FP8 quantization (~115–130GB VRAM for single GPU; verify on your hardware). MoE activation reduces per-token compute but does not reduce peak memory footprint. Multi-GPU inference recommended for production. Known distributed inference issues with transformers fine-grained FP8; vLLM/SGLang preferable. Requires CUDA 11.8+, modern GPU (A100, H100, or equivalent).

Card does not explicitly document LoRA, QLoRA, or full fine-tuning feasibility. MoE models typically require careful expert routing updates. Recommend consulting official Qwen GitHub/documentation before fine-tuning. FP8 quantization may complicate gradient-based tuning; full precision (BF16) checkpoint may be required for training.

When to avoid it — and what to weigh

  • Latency-Critical, Single-Token Applications — Despite MoE efficiency, 235B base model incurs significant computational overhead. Non-thinking mode is faster but still requires substantial resources for sub-100ms response budgets.
  • Limited GPU Memory (<80GB VRAM) — FP8 quantization reduces footprint, but full model still requires substantial VRAM. Distributed inference on transformers has known fine-grained FP8 issues; consider vLLM/SGLang instead.
  • Proprietary, Closed-Ecosystem Deployment — Model is open-source, but orchestration via SGLang/vLLM adds operational complexity. Requires careful version pinning and CUDA_LAUNCH_BLOCKING workarounds for multi-device setups.
  • Real-Time Streaming with Thinking Mode — Thinking mode generates variable-length internal reasoning before output, making token-streaming unpredictable. Non-thinking mode better suited for streaming APIs.

License & commercial use

Apache 2.0 license. Permissive OSI license; permits commercial use, modification, and distribution with attribution and liability notice included.

Apache 2.0 is a permissive, OSI-approved license permitting commercial deployment. No usage restrictions, licensing fees, or commercial use caveats stated. However, Alibaba/Qwen retains no indemnification for model output. Verify model output compliance with your jurisdiction's AI/LLM regulations (e.g., EU AI Act, local data residency).

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

No security audit, vulnerability disclosure, or adversarial robustness claims stated. Model is open-source; source code and weights publicly available for review. Typical LLM risks apply: prompt injection, jailbreaking, hallucinations, bias. Thinking mode generates visible internal reasoning, which may aid interpretability but also expose model reasoning to manipulation. FP8 quantization may alter inference behavior; test on target hardware. No claims of differential privacy, watermarking, or safety alignment beyond general instruction-following.

Alternatives to consider

QwQ-32B or Qwen2.5-Instruct

Smaller alternatives from Qwen series. QwQ-32B pure reasoning; Qwen2.5-Instruct pure instruction-following. Choose if memory-constrained or if thinking/non-thinking switching unnecessary.

Llama 3.1 70B / 405B

Dense transformer, no MoE. Broader ecosystem support (vLLM, TGI, llama.cpp). Smaller models (70B) lower memory; larger (405B) higher performance. No native thinking mode.

Mistral Large or Mixtral 8x22B

Smaller MoE alternatives. Mistral Large is dense and well-supported; Mixtral is MoE with 22B active. Simpler architecture, mature tooling; lower reasoning capability claim than Qwen3.

Software development agency

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

Review hardware requirements, set up vLLM or SGLang for production inference, and test thinking/non-thinking modes on your workload. Contact our team for architecture review and cost estimation.

Talk to DEV.co

Related open-source tools

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

Qwen3-235B-A22B-FP8 FAQ

Can I use this model commercially?
Yes. Apache 2.0 is permissive and allows commercial use without licensing fees. However, you are responsible for model outputs (hallucinations, bias, regulatory compliance) and must comply with your local AI/data regulations.
What is the minimum GPU memory required?
Approximately 115–130GB VRAM for FP8 single-GPU inference. Use vLLM or SGLang for multi-GPU setups. Distributed inference on transformers has known issues; consider alternatives if memory is critical.
What does 'thinking mode' do, and when should I use it?
Thinking mode enables the model to generate explicit internal reasoning (wrapped in <think>...</think>) before responding. Use for complex math, code, or logic tasks. Non-thinking mode is faster and suits general dialogue. Toggle via enable_thinking=True/False in chat template.
Does FP8 quantization affect accuracy?
Card claims fine-grained FP8 maintains quality comparable to BF16, but this is model-specific. Test on your benchmark. FP8 distributed inference in transformers has reported issues; verify with vLLM/SGLang on multi-GPU setups.

Software development & web development with DEV.co

DEV.co helps companies turn open-source tools like Qwen3-235B-A22B-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.

Ready to Deploy Qwen3?

Review hardware requirements, set up vLLM or SGLang for production inference, and test thinking/non-thinking modes on your workload. Contact our team for architecture review and cost estimation.