DEV.co
Open-Source LLM · Qwen

Qwen3-8B-FP8

Qwen3-8B-FP8 is an 8.2B-parameter open-source language model from Alibaba's Qwen team. It offers dual modes: a 'thinking' mode for complex reasoning (math, code, logic) and a 'non-thinking' mode for fast general dialogue. The model supports 100+ languages and can handle 32K tokens natively, extending to 131K with YaRN. This FP8-quantized version trades some precision for reduced memory footprint and faster inference while maintaining compatibility with standard inference frameworks.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-8B-FP8
8.2B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
626.5k
Downloads (30d)

Key facts

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

FieldValue
DeveloperQwen
Parameters8.2B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads626.5k
Likes62
Last updated2025-07-26
SourceQwen/Qwen3-8B-FP8

What Qwen3-8B-FP8 is

Qwen3-8B is a causal language model with 36 layers, 32 query heads and 8 key-value heads (GQA), trained with both pretraining and post-training stages. The FP8 variant uses fine-grained block-wise quantization (block size 128) and is distributed as safetensors. Native context is 32,768 tokens; YaRN extends to 131,072. The model includes a soft-switch mechanism (/think, /no_think) for dynamic mode toggling within conversations. Requires transformers ≥4.51.0.

Quickstart

Run Qwen3-8B-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/Qwen3-8B-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 problem-solving and technical Q&A

Leverage thinking mode for math, code generation, and multi-step reasoning where intermediate work improves accuracy. Soft-switch allows toggling per question without reloading.

Multilingual conversational AI and customer support

Native support for 100+ languages with strong instruction-following makes it suitable for global applications. Non-thinking mode ensures low latency for real-time chat.

Agentic workflows and tool integration

Demonstrated agent capabilities in both modes allow orchestration of external APIs and services. Suitable for autonomous task execution with reasoning transparency.

Running & fine-tuning it

ESTIMATE: FP8 quantization reduces memory vs. bfloat16. Typical GPU VRAM for inference: 16–24 GB (single A100/L40S) for batch inference, 8–12 GB for single-turn chat. Exact requirements depend on context length, batch size, and inference framework. See Qwen blog/docs for benchmarks. Multi-GPU setups require vLLM or SGLang to avoid fine-grained FP8 transformers bug.

Not explicitly stated in card. No LoRA, QLoRA, or instruction-tuning guidance provided. Infer: base model is post-trained; further fine-tuning feasibility unknown. Recommend checking Qwen GitHub or documentation for adapter strategies.

When to avoid it — and what to weigh

  • Extreme latency constraints without thinking enabled — Thinking mode generates intermediate reasoning tokens before responses, increasing total latency. Non-thinking mode is faster but sacrifices reasoning depth. Check latency requirements first.
  • Greedy decoding workflows — Model card explicitly warns against greedy decoding in thinking mode, which causes performance degradation and infinite loops. Requires sampling (Temperature 0.6, TopP 0.95, etc.).
  • Multi-GPU distributed inference with fine-grained FP8 on transformers — Known issue: fine-grained FP8 in transformers has problems under distributed inference. Workaround is CUDA_LAUNCH_BLOCKING=1, but vLLM or SGLang are recommended.
  • Applications requiring guaranteed output length or streaming truncation safety — No explicit documentation on output length guarantees, streaming truncation behavior, or edge cases in thinking-mode token counting. Requires testing.

License & commercial use

Apache-2.0 license. Permissive OSI-approved open-source license allowing broad use, modification, and redistribution.

Apache-2.0 is a permissive license compatible with commercial use. No gating restrictions (gated=false). However, consult your legal team and Alibaba Qwen's terms for any proprietary limitations not encoded in the license file, especially regarding model output attribution and derivatives. No commercial support vendor noted in card.

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 explicit security audit, adversarial robustness, or prompt-injection mitigations stated. As with all LLMs, model outputs can be manipulated by prompt engineering and may reflect training data biases. Thinking mode adds complexity (token parsing, special tokens like <think></think>); validate thinking-content extraction logic. No mention of content filtering or jailbreak defenses. Requires your own safety evaluation.

Alternatives to consider

Qwen2.5-Instruct

Smaller/faster option if reasoning depth not needed; Qwen3 cited as superior on math/code/reasoning. If full multi-mode flexibility unnecessary, older version may reduce resource overhead.

Meta Llama 3.1 or 3.2 (8B)

Comparable 8B scale, Apache-2.0 license, strong open-source ecosystem. No built-in thinking mode; trade reasoning transparency for simplicity and broader tool support.

Deepseek-R1 or QwQ-32B (larger sibling)

If advanced reasoning is critical, larger models offer deeper chain-of-thought. Deepseek-R1 and QwQ are reasoning-focused; trade size and cost for better complex task performance.

Software development agency

Ship Qwen3-8B-FP8 with senior software developers

Explore self-hosted LLM deployment, custom LLM applications, and RAG integrations. Start with vLLM or SGLang for production, or test locally with Ollama. Review Qwen GitHub and documentation for benchmarks and fine-tuning.

Talk to DEV.co

Related open-source tools

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

Related on DEV.co

Explore the category and the services that help you build with it.

Qwen3-8B-FP8 FAQ

Can I use Qwen3-8B-FP8 commercially without additional licensing?
Yes, Apache-2.0 is permissive. However, confirm with Alibaba/Qwen and your legal counsel for any additional terms, trademark, or attribution requirements outside the license itself.
What GPU VRAM do I need to run this model?
Estimate 8–24 GB depending on context, batch size, and inference framework. FP8 quantization reduces memory vs. bfloat16. Use vLLM or SGLang for production to avoid transformers fine-grained FP8 bugs on multi-GPU setups.
How does thinking mode affect response latency?
Thinking mode generates hidden reasoning before the final answer, increasing total latency. Use non-thinking mode (enable_thinking=False) for speed-critical applications. Soft-switch (/think, /no_think) allows per-turn control.
Is there a quantized version smaller than FP8?
Not stated in card. Check Qwen GitHub/HuggingFace for INT4/INT8/GGUF variants. Only FP8 is documented here.

Software development & web development with DEV.co

DEV.co helps companies turn open-source tools like Qwen3-8B-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 Qwen3-8B-FP8?

Explore self-hosted LLM deployment, custom LLM applications, and RAG integrations. Start with vLLM or SGLang for production, or test locally with Ollama. Review Qwen GitHub and documentation for benchmarks and fine-tuning.