Qwen3 30B-A3B: The Original Release and Its 2507 Refresh

What Qwen3 30B-A3B is

Qwen3 30B-A3B is Alibaba's small mixture-of-experts model: 30.5 billion total parameters with 3.3 billion active per token, Apache 2.0 and self-hostable, with a native context of 32,768 tokens extended to 131,072 with YaRN, at roughly $0.13 per million input tokens and $0.52 output through third-party providers. It is the original release, and Alibaba later published Qwen3-30B-A3B-Instruct-2507 from the same architecture, covered separately here.

Long-context results against the refresh

Alibaba publish accuracy for both versions on a 1M-token RULER benchmark, at every length from 4K to 1,000K. In non-thinking mode the original averages 72.0%. It starts strong at 97.1% on 4K inputs and holds 92.2% at 32K, then 76.9% at 128K, 61.9% at 384K, and 50.9% at a million tokens. The 2507 refresh, same size and same architecture, averages 86.8% and holds 72.8% at a million.

LMArena

On LMArena's text leaderboard Qwen3 30B A3b has 1327 Elo (rank 253); by category, 1385 Elo in Coding (rank 234), 1351 Elo in Math (rank 219), 1344 Elo in Expert (rank 231) and 1283 Elo in Creative Writing (rank 264). The 2507 refresh has 1382 Elo (rank 191).

Thinking and non-thinking modes

Unlike the later 2507 releases, this model switches between a thinking mode for reasoning, mathematics and coding and a non-thinking mode for general dialogue within one model. Thinking is on by default and can be disabled with enable_thinking=False in the chat template; with thinking enabled, users can also add /think or /no_think to a prompt or system message to switch from turn to turn. Alibaba recommend temperature 0.6, top_p 0.95, top_k 20 and min_p 0 for thinking mode, and warn against greedy decoding, which can cause repetition; for non-thinking mode they suggest temperature 0.7, top_p 0.8, top_k 20 and min_p 0.

Specifications

The card lists 29.9 billion non-embedding parameters, 48 layers, 32 query heads and 4 key-value heads, and 128 experts with 8 activated. Alibaba describe Qwen3 as supporting more than 100 languages and dialects. It needs Transformers 4.51.0 or later and can be served with SGLang 0.4.6.post1 or vLLM 0.8.5, and Ollama, LM Studio, MLX-LM, llama.cpp and KTransformers support it locally.

Long contexts and output length

Qwen3 natively supports 32,768 tokens. Alibaba validated performance up to 131,072 tokens with YaRN, which Transformers, llama.cpp, vLLM and SGLang support, and note that static YaRN can affect performance on shorter texts, so it should be enabled only when needed. Alibaba's example generates up to 32,768 new tokens. For deployment, SGLang uses the qwen3 reasoning parser and vLLM the deepseek_r1 parser. Alibaba state that thinking mode surpasses QwQ and that non-thinking mode surpasses the Qwen2.5 instruct models on mathematics, code generation and commonsense reasoning, and point to Qwen-Agent for tool calling.

Lineage

Qwen3 is a generation of dense and mixture-of-experts models, and Qwen3-30B-A3B is its smaller MoE model. Alibaba describe its tool-calling as working in both thinking and non-thinking modes, and recommend Qwen-Agent, which wraps tool-calling templates and parsers and accepts MCP server configuration files.

Local use

Alibaba list Ollama, LM Studio, MLX-LM, llama.cpp and KTransformers as supporting Qwen3 for local use, and the model can also be called through Alibaba Model Studio.

Official Resources