DEV.co
Open-Source LLM · Qwen

Qwen3-30B-A3B-Thinking-2507-FP8

Qwen3-30B-A3B-Thinking-2507-FP8 is a 30.5B parameter open-source language model from Alibaba's Qwen team, released July 2025. It is a mixture-of-experts (MoE) model with 8 active experts per token, optimized for reasoning tasks (mathematics, logic, coding) and long-context understanding (up to 256K tokens). The FP8 quantization reduces memory footprint while maintaining inference quality. It is gated=false, permissively licensed (Apache 2.0), and compatible with standard deployment frameworks (vLLM, SGLang, Ollama).

Source: HuggingFace — huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507-FP8
30.5B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
45.1k
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
Downloads45.1k
Likes67
Last updated2025-07-30
SourceQwen/Qwen3-30B-A3B-Thinking-2507-FP8

What Qwen3-30B-A3B-Thinking-2507-FP8 is

30.5B total parameters (3.3B activated per token), 48 layers, 32 query heads + 4 KV heads (GQA), 128 experts with 8 activated. Native context length 262,144 tokens. MoE architecture reduces compute vs. dense models. FP8 quantization uses block size 128. Implements a thinking/reasoning mode (automatic, no flag required). Requires transformers>=4.51.0. Latest model card dated 2025-07-30. Supports tool calling and agentic workflows via Qwen-Agent.

Quickstart

Run Qwen3-30B-A3B-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/Qwen3-30B-A3B-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

Complex reasoning and mathematics

AIME25 score 85.0 and HMMT25 score 71.4 indicate strong performance on advanced math and competition-style problems. Extended thinking overhead allows deep reasoning chains.

Autonomous agents and tool use

Excels in BFCL-v3 (72.4) and TAU benchmarks; native tool calling support via chat template. Qwen-Agent library simplifies agentic deployment.

Cost-sensitive reasoning at scale

FP8 quantization + MoE activation (only 3.3B of 30.5B per token) reduces GPU memory and inference latency vs. dense 30B models, enabling cost-effective batch processing.

Running & fine-tuning it

FP8 quantization: estimated ~28–35 GB VRAM for single-GPU inference (30.5B params + KV cache overhead at 262K context). Recommended: 2x H100 (80GB) or equivalent for throughput. BF16 unquantized: ~60+ GB VRAM. Context length >131K strongly recommended for reasoning tasks per documentation.

Not explicitly documented in model card. MoE models (with multiple expert modules) typically require careful learning-rate tuning and may have convergence challenges. QLoRA feasibility depends on peak memory usage during backprop; recommend experimental validation. No official LoRA adapters or fine-tuning templates provided.

When to avoid it — and what to weigh

  • Sub-millisecond latency requirements — MoE models incur expert routing overhead. Reasoning mode generates lengthy internal thought tokens, increasing first-token latency. Not suitable for real-time user-facing chat.
  • Single-GPU (consumer-grade) deployment without quantization — 30.5B model requires significant VRAM. FP8 helps but running unquantized BF16 on a single 40GB GPU is marginal. Recommend multi-GPU or further quantization (INT8/INT4).
  • Domain-specific fine-tuning at scale — No mention of LoRA/QLoRA guidance in card. MoE models have complex gradient dynamics; fine-tuning feasibility and cost not documented. Requires testing.
  • Production security-critical applications without external audit — Model has not undergone third-party security or bias audits (not stated in card). Reasoning chains may generate plausible but unverified information; requires human-in-the-loop validation.

License & commercial use

Apache License 2.0 (apache-2.0). Permissive OSI-approved license allowing commercial and derivative use with attribution and notice of modifications. No usage restrictions for training, inference, or deployment.

Apache 2.0 explicitly permits commercial use, including serving as a SaaS, re-selling, and derivative models. No gating, no paid tier required. End-user applications, API wrappers, and proprietary fine-tunes are legally permitted. No warranty or liability clause in Apache 2.0; standard risk applies to any LLM output (hallucinations, accuracy, bias).

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 risks: model may generate plausible but unverified outputs (hallucination). Thinking chains are opaque (explainability limited). No adversarial robustness or jailbreak evaluations published. Input/output filtering responsibility rests with the deployer. FP8 quantization may marginally reduce attack surface vs. higher precision, but security is not a stated design goal. Recommend input validation, content moderation, and human review for high-stakes use.

Alternatives to consider

Qwen3-235B-A22B-Thinking

Larger Qwen model (235B) with higher absolute performance on some MMLU and reasoning tasks but requires proportionally more compute. Choose if budget/latency permit and max performance is critical.

GPT-4-Turbo with extended thinking (proprietary)

Closed-source, higher cost, but established production track record and commercial SLA support. Suitable if vendor lock-in and higher cost are acceptable trade-offs for enterprise guarantees.

Deepseek-R1 (distilled variants)

Open-source reasoning model; vLLM/SGLang reference Deepseek reasoning parser. Smaller variants may offer faster inference, but full benchmark comparison to Qwen3-30B not provided in this card.

Software development agency

Ship Qwen3-30B-A3B-Thinking-2507-FP8 with senior software developers

Ready to run advanced reasoning workloads? Start with vLLM or SGLang on our GPU infrastructure, or download and self-host via Hugging Face. Apache 2.0 licensed—no restrictions. Contact our team for custom deployment, fine-tuning guidance, or multi-GPU orchestration.

Talk to DEV.co

Related open-source tools

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

Qwen3-30B-A3B-Thinking-2507-FP8 FAQ

Can I use this model commercially?
Yes. Apache 2.0 license explicitly permits commercial use, including SaaS, fine-tuning for proprietary products, and re-distribution. No paid license or approval required. Standard model-output disclaimers (hallucination, accuracy) apply.
How much VRAM do I need?
FP8 quantization: ~28–35 GB for single-GPU inference (including KV cache). For 256K context, allocate higher. BF16 unquantized: ~60+ GB. Multi-GPU recommended for production workloads. Test with your target hardware.
What is 'thinking mode' and is it mandatory?
Thinking mode is enabled by default and cannot be disabled. The model automatically includes internal reasoning tokens (between <think> and </think> tags). Output length increases; use the provided code snippet to parse thinking vs. final answer. Recommended for complex reasoning tasks.
Can I fine-tune this model?
Not explicitly covered in the model card. MoE models require careful tuning. QLoRA may be possible but not tested by Qwen team. No official LoRA adapters provided. Experimental validation on your hardware/dataset is recommended before committing to fine-tuning.

Work with a software development agency

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Qwen3-30B-A3B-Thinking-2507-FP8 is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Deploy Qwen3-30B-A3B-Thinking-2507-FP8 Today

Ready to run advanced reasoning workloads? Start with vLLM or SGLang on our GPU infrastructure, or download and self-host via Hugging Face. Apache 2.0 licensed—no restrictions. Contact our team for custom deployment, fine-tuning guidance, or multi-GPU orchestration.