Qwen3-30B-A3B

Qwen3-30B-A3B is the mixture-of-experts entry in Alibaba's Qwen 3 family, built to prove a point: a model that only turns on 3.3 billion of its 30.5 billion parameters per token can still compete with dense models several times its active size. Released April 29, 2025 alongside the rest of the Qwen3 line, it shipped with a hybrid thinking mode that lets a single checkpoint switch between step-by-step reasoning and fast direct answers depending on the prompt.

What makes the model worth tracking a year after launch is the update cycle around it. Alibaba refreshed it in July 2025 with separate Instruct-2507 and Thinking-2507 checkpoints, then spun off a dedicated Qwen3-Coder-30B-A3B-Instruct variant for agentic coding. Each kept the same 30.5B/3.3B architecture while retraining for a specific job, which is the opposite of the "one model does everything" pitch most labs make.

The model's efficiency claim also gave it an odd second life: independent researchers have used it as a benchmark subject for everything from consumer-GPU inference studies to cross-lingual safety evaluations, exactly because it's small enough to run on a single high-end GPU and open enough to instrument freely.

Key Specifications

The 128-expert, 8-active routing ratio is the core efficiency story. Only about 6% of total parameters fire on any given token, which is why Alibaba's own comparison calls out that the model "outcompetes QwQ-32B with 10 times of activated parameters" - QwQ-32B being the prior-generation reasoning model that needed its full 32B active to hit similar scores.

Benchmark Performance

Alibaba's July 2025 refresh makes for the clearest before-and-after picture. The table below uses Qwen's own published numbers comparing the original non-thinking mode against the Instruct-2507 update, the larger Qwen3-235B-A22B flagship of the era, and two proprietary references.

The jump between the original release and the 2507 update isn't a rounding error. AIME25 nearly triples (21.6 to 61.3) and Arena-Hard v2 almost triples (24.8 to 69.0), which tells you the original post-training left real capability on the table that better data and tuning recovered without touching the architecture. Even after that jump, the 235B-A22B flagship still leads on knowledge-heavy benchmarks like MMLU-Pro and GPQA, since raw parameter count still matters for breadth of world knowledge. On agentic tool use (BFCL-v3) and instruction following (IFEval), the gap between the 30B-A3B and its much larger sibling nearly closes.

At 3.3B active parameters, quantized builds of Qwen3-30B-A3B fit comfortably on a single high-end consumer GPU. Source: unsplash.com

Key Capabilities

The defining feature at launch was the hybrid thinking toggle: the same weights could run in enable_thinking=True mode for math and multi-step logic, or skip the <think> block completely for fast conversational replies. Alibaba later split this into two dedicated checkpoints instead - Thinking-2507, which pushes AIME25 to 85.0 in exchange for verbose reasoning traces, and Instruct-2507, the non-thinking model behind the benchmark table above. If you need one model that does both well without picking a mode, the original hybrid checkpoint is still the only option in the family that has it.

Agentic tool use is a genuine strength for the size class. BFCL-v3 scores in the mid-60s put it within a few points of GPT-4o, and the Qwen3-Coder-30B-A3B-Instruct variant extends that further into repository-scale coding and agentic browser use, following the pattern set by the larger Qwen3-Coder-Next. None of the three 30B-A3B checkpoints changed the underlying MoE architecture; they differ purely in post-training data and objective.

Multilingual coverage across 119 languages and dialects is broad enough that the model shows up in cross-lingual research well beyond typical chatbot benchmarking - including recent work using it to test how model behavior shifts when prompted in different languages.

Qwen3-30B-A3B supports 119 languages and dialects, a range wide enough that researchers use it in cross-lingual behavioral studies. Source: unsplash.com

Pricing and Availability

Qwen3-30B-A3B is Apache 2.0 licensed with no usage restrictions, and the weights for all four checkpoints (base, Instruct-2507, Thinking-2507, Coder-30B-A3B-Instruct) are downloadable from Hugging Face, ModelScope, and Kaggle, with GGUF, MLX, and FP8 quantizations available from community maintainers. Our self-hosting guide puts the Q4 download at roughly 19GB with a 20GB minimum VRAM footprint - it fits on a single RTX 4090 with room to spare, and runs at a fraction of the latency of the dense Qwen3-32B on the same card.

For teams that would rather not manage inference infrastructure, hosted API pricing varies noticeably by provider:

That's a 7x spread between the cheapest and most expensive host for what is, underneath, identical weights - worth checking before committing to a provider for anything beyond hobby use. Recommended sampling settings from Qwen are temperature 0.7, top-p 0.8, top-k 20, with vLLM (0.8.5+) or SGLang (0.4.6.post1+) for production serving.

Strengths

• 3.3B active parameters deliver benchmark scores that rival much larger dense models, especially after the 2507 update
• Apache 2.0 license with no restrictions on commercial use or fine-tuning
• Runs on a single consumer GPU at Q4 quantization - genuinely practical for local deployment
• 119-language coverage is unusually broad for this parameter class
• Specialized siblings (Thinking-2507, Coder-30B-A3B-Instruct) let you pick a checkpoint tuned for your actual workload

Weaknesses

• The original April 2025 release underperformed its own architecture's ceiling - AIME25 nearly tripled after the July retraining, meaning early adopters were judging the checkpoint on numbers that didn't hold up
• Scored 71.43% max harm rate in independent multi-turn jailbreak testing, rated "highly vulnerable" with Gemini 2.5 Flash
• Trails the 235B-A22B flagship on pure knowledge benchmarks (MMLU-Pro, GPQA) where total parameter count still matters
• Hosted API pricing varies up to 7x across providers for identical weights, so the "free" license doesn't guarantee cheap inference
• Now one generation behind the newer Qwen3.5-35B-A3B, which surpasses even the 235B-A22B flagship this model trails on knowledge tasks

Related Coverage

• Qwen 3 Review - our review of the full Qwen 3 family this model belongs to
• Qwen3.5-35B-A3B - the successor MoE model that leapfrogs Qwen3-30B-A3B on nearly every benchmark
• Qwen3-Coder-Next - the larger, coding-focused MoE sibling
• Best Open-Source LLMs You Can Self-Host in 2026 - VRAM and quantization data for running this model locally
• Open Source LLM Leaderboard
• Small Language Model Leaderboard
• AI Models Can Now Jailbreak Other AI Models Autonomously - jailbreak-resistance testing that flagged this model as highly vulnerable
• Liquid AI's LFM2-24B benchmarks itself against Qwen3-30B-A3B as its primary efficiency target

Sources:

• Qwen3-30B-A3B Model Card (Hugging Face)
• Qwen3-30B-A3B-Instruct-2507 Model Card (Hugging Face)
• Qwen3-30B-A3B-Thinking-2507 Model Card (Hugging Face)
• Qwen3-Coder-30B-A3B-Instruct Model Card (Hugging Face)
• Qwen3: Think Deeper, Act Faster (official blog)
• Qwen3 Technical Report, An Yang et al., arXiv:2505.09388
• QwenLM/Qwen3 GitHub Repository
• Qwen3 30B A3B - API Pricing & Benchmarks (OpenRouter)