DEV.co
Open-Source LLM · Qwen

Qwen3-30B-A3B-FP8

Qwen3-30B-A3B-FP8 is a 30.5B-parameter mixture-of-experts (MoE) language model from Alibaba's Qwen team, with only 3.3B parameters active per token. It supports dual-mode operation: a thinking mode for complex reasoning (math, code, logic) and a non-thinking mode for fast general dialogue. The model is quantized to FP8 for reduced memory footprint, supports 32K native context (up to 131K with YaRN), and is trained on 100+ languages. It is available under Apache 2.0 and not gated.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-30B-A3B-FP8
30.5B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
450k
Downloads (30d)

Key facts

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

FieldValue
DeveloperQwen
Parameters30.5B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads450k
Likes84
Last updated2025-07-26
SourceQwen/Qwen3-30B-A3B-FP8

What Qwen3-30B-A3B-FP8 is

FP8-quantized MoE causal language model with 48 transformer layers, 128 total experts (8 active per token), 32 query heads + 4 KV heads (GQA), and 29.9B non-embedding parameters. Native context window 32,768 tokens; supports extension to 131,072 via YaRN. Compatible with transformers>=4.51.0, vLLM>=0.8.5, SGLang>=0.4.6.post1, and local inference via Ollama, LMStudio, MLX-LM, llama.cpp, KTransformers. Requires fine-grained FP8 quantization support; distributed inference on transformers known to have issues with FP8 (may require CUDA_LAUNCH_BLOCKING=1).

Quickstart

Run Qwen3-30B-A3B-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/Qwen3-30B-A3B-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 and code generation with flexible efficiency trade-off

Use thinking mode for math, logic puzzles, and code generation where reasoning steps improve quality. Switch to non-thinking mode for the same tasks when latency/cost is critical. Single model eliminates need to maintain separate reasoning and fast-inference variants.

Multilingual chatbot and agent applications

Strong instruction-following, creative writing, and role-play capabilities across 100+ languages. Agent mode enables precise tool integration. Suitable for customer support, content generation, and multi-turn dialogue systems.

Cost-efficient self-hosted deployment on limited GPU memory

FP8 quantization + MoE sparsity (8 of 128 experts active) reduces memory footprint significantly. Deployable on single consumer/prosumer GPUs. SGLang and vLLM support enable OpenAI-compatible API endpoints without external inference services.

Running & fine-tuning it

ESTIMATE: FP8 quantization (~31GB model size / 8 bits per param, plus overhead) suggests ~15–20GB GPU VRAM for inference on single GPU. MoE sparsity reduces per-token compute but not memory footprint. Exact requirements depend on batch size, context length, and serving framework. Multi-GPU setups or CPU+GPU offloading may be needed for thinking mode with long contexts. Verify with your inference framework's memory profiler.

Card does not explicitly document LoRA or QLoRA support for this quantized variant. Base model (Qwen3-30B-A3B) is post-trained; fine-tuning typically applied to unquantized BF16 checkpoint. FP8-quantized weights may require specialized quantization-aware fine-tuning or retraining of quantization parameters. Requires review of Qwen3 GitHub/documentation for official LoRA/QLoRA guidance. Recommend contacting Qwen team or consulting their GitHub issues for fine-tuning best practices on FP8 variant.

When to avoid it — and what to weigh

  • Real-time ultra-low-latency applications requiring <100ms inference — Even with thinking mode disabled and FP8 quantization, 30B+ parameters will incur measurable latency. Thinking mode adds further overhead. Evaluate with your specific hardware; consider smaller 8B or 14B models if latency is critical.
  • Limited context requirements but strict determinism needed — Thinking mode uses sampled decoding (Temperature=0.6, TopP=0.95) to avoid performance degradation and endless loops. If you require greedy/deterministic output, disable thinking mode, but you lose the reasoning advantage. Non-thinking mode still uses sampled decoding (Temperature=0.7).
  • Edge deployment or CPU-only environments — Model size and FP8 quantization design target GPU inference. Quantization support for edge runtimes (llama.cpp, MLX) is available but not officially benchmarked. Local inference frameworks may lag behind official support.
  • Distributed multi-GPU inference with transformers library — Known issue: fine-grained FP8 in transformers has compatibility issues under distributed setup. Requires workaround (CUDA_LAUNCH_BLOCKING=1) or use vLLM/SGLang instead.

License & commercial use

Apache 2.0 (OSI-approved, permissive open-source license). Allows commercial use, modification, and redistribution with minimal obligations: retain license and copyright notice, and disclose changes.

Apache 2.0 is a permissive OSI license with no restrictions on commercial use. You may use, modify, and redistribute Qwen3-30B-A3B-FP8 in proprietary products provided you: (1) include Apache 2.0 license text in distribution, (2) disclose material modifications, (3) state changes made to the code. No patent grant, liability limitations, or trademark use restrictions are provided beyond standard Apache 2.0 terms. This is clearer than many LLM licenses (e.g., Llama 2 Community License), but consult legal counsel for enterprise deployments to confirm compliance with your specific use case.

DEV.co evaluation signals

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

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

No explicit security or safety features documented in card. Standard LLM considerations apply: (1) Model may generate harmful, biased, or false content; implement content filtering or human review in production. (2) Thinking mode outputs internal reasoning tokens; ensure logs and API endpoints do not expose sensitive user data or model internals inadvertently. (3) As an open model, weights are public; no confidentiality of model behavior. (4) FP8 quantization may introduce rounding artifacts; not suitable for cryptographic or security-critical applications. (5) Recommend security testing and prompt injection evaluation before production deployment, especially in agent-based workflows with tool integration.

Alternatives to consider

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

Qwen's previous-generation dense models. If you do not need dual thinking/non-thinking modes or MoE sparsity, Qwen2.5 may be simpler and have more mature deployment tooling. Trade-off: no reasoning mode, larger parameter count for same performance tier.

DeepSeek-R1 or similar open reasoning models

Alternative reasoning-focused open models. DeepSeek-R1 offers similar thinking/non-thinking capability. Evaluate on benchmarks (math, code, reasoning) specific to your use case. Note: DeepSeek uses different license (MIT or proprietary); verify commercial terms.

Mixtral 8x7B or similar open MoE models

If MoE efficiency is primary goal but reasoning is secondary, Mixtral or similar open MoE may be smaller and faster. Trade-off: no dual reasoning mode, less multilingual support, older training data. Suitable for cost-constrained scenarios.

Software development agency

Ship Qwen3-30B-A3B-FP8 with senior software developers

Start with vLLM or SGLang for production-grade inference, or experiment locally with Ollama or llama.cpp. Review hardware requirements, test thinking vs. non-thinking trade-offs with your workload, and ensure Apache 2.0 compliance for commercial use. Contact Devco for guidance on infrastructure, fine-tuning, or production deployment.

Talk to DEV.co

Related open-source tools

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

Qwen3-30B-A3B-FP8 FAQ

Can I use Qwen3-30B-A3B-FP8 in a commercial product without paying Qwen?
Yes. Apache 2.0 permits commercial use without royalties or licensing fees. You must: (1) include the Apache 2.0 license and copyright notice in your product, (2) disclose any material changes you make to the model weights or code, (3) state that the model is derived from Qwen. Consult your legal team for compliance with any proprietary data or usage policies your business may have.
What GPU do I need to run this model?
Estimated 15–20GB VRAM for single-GPU inference (FP8 quantization). A100 (40GB), H100 (80GB), or RTX 6000 (48GB) would be comfortable. Smaller GPUs (RTX 4090, RTX 3090) at 24GB may work with batch size=1 and careful memory management. Multi-GPU or CPU offloading required for larger batches or long contexts. Use vLLM or SGLang profiling to measure exact requirements on your hardware.
How much slower is thinking mode than non-thinking mode?
Card does not provide latency benchmarks. Thinking mode generates internal reasoning tokens before the final response, adding latency proportional to reasoning complexity. For simple queries, overhead may be 2–5x; for hard reasoning, offset by better quality. Test with your workload using vLLM or SGLang to measure end-to-end latency and throughput.
Can I fine-tune the FP8 model directly?
Not documented. Fine-tuning typically requires unquantized BF16 weights. FP8 quantization may require quantization-aware fine-tuning or re-quantization afterward. Check Qwen GitHub and readthedocs for official LoRA/QLoRA guidance. Contact Qwen team if fine-tuning the FP8 checkpoint is your requirement.

Software development & web development with DEV.co

Need help beyond evaluating Qwen3-30B-A3B-FP8? 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 Qwen3-30B-A3B-FP8?

Start with vLLM or SGLang for production-grade inference, or experiment locally with Ollama or llama.cpp. Review hardware requirements, test thinking vs. non-thinking trade-offs with your workload, and ensure Apache 2.0 compliance for commercial use. Contact Devco for guidance on infrastructure, fine-tuning, or production deployment.