Qwen3-30B-A3B-Instruct-2507: 256K Context and RULER Results

What this model is

Qwen3-30B-A3B-Instruct-2507 is the July 2025 refresh of Alibaba's 30 billion parameter mixture-of-experts model, with 3.3 billion parameters active per token (30.5 billion in total), Apache 2.0 and self-hostable, carrying a 256K token context window. It is roughly $0.05 per million input tokens and $0.19 output through third-party providers.

Long-context results, original and refresh

Alibaba publish accuracy on a 1M-token RULER benchmark at every length from 4K to 1,000K, for the original and the refresh side by side. The original Qwen3-30B-A3B in non-thinking mode averages 72.0%: it starts at 97.1% on 4K inputs, is still respectable at 92.2% by 32K, then 76.9% at 128K, 61.9% at 384K, and 50.9% at a million tokens. The 2507 refresh averages 86.8%, holding 97.2% at 32K, 89.1% at 128K, 83.6% at 384K and 72.8% at 1,000K. Sparse attention matches full attention exactly on the average, at 86.8%, and tracks it closely across the range.

Full and sparse attention

Alibaba measured the 2507 refresh under both full and sparse attention; the averages are identical at 86.8%. Sparse is fractionally ahead at 64K, 96K and 384K, and fractionally behind at 32K and 128K.

LMArena and prompting

On LMArena's text leaderboard Qwen3 30B A3b Instruct 2507 has 1382 Elo (rank 191); by category, 1437 Elo in Coding (rank 173), 1378 Elo in Math (rank 194), 1396 Elo in Expert (rank 189) and 1318 Elo in Creative Writing (rank 217). Alibaba document that mathematics prompts should include the instruction to reason step by step and put the final answer in a box.

Specifications

The card lists 29.9 billion non-embedding parameters, 48 layers, grouped-query attention with 32 query heads and 4 key-value heads, and 128 experts with 8 activated, with a native context of 262,144 tokens. It supports only non-thinking mode. Running it at about 1 million tokens requires approximately 240 GB of total GPU memory, per Alibaba, using tensor parallelism of 4.

Alibaba’s comparison table

Against the original Qwen3-30B-A3B in non-thinking mode, Alibaba report MMLU-Pro 78.4 against 69.1, GPQA 70.4 against 54.8, AIME25 61.3 against 21.6, ZebraLogic 90.0 against 33.2, LiveCodeBench v6 43.2 against 29.0, Arena-Hard v2 69.0 against 24.8 and WritingBench 85.5 against 72.2. In the same table Gemini 2.5 Flash without thinking records 81.1 on MMLU-Pro and 61.6 on AIME25.

Agentic and multilingual results

On Alibaba's agent and multilingual tests it records BFCL-v3 65.1, TAU1-Retail 59.1, TAU2-Retail 57.0, TAU2-Telecom 12.3, MultiIF 67.9, MMLU-ProX 72.0 and PolyMATH 43.1. The original non-thinking release scored 58.6, 38.3, 31.6, 18.4, 70.8, 65.1 and 23.3. Arena-Hard v2 results are win rates judged by GPT-4.1. The model needs Transformers 4.51.0 or later; Alibaba's examples serve it with SGLang 0.4.6.post1 or vLLM 0.8.5 and advise reducing the context to 32,768 tokens if memory runs out. Ollama, LM Studio, MLX-LM, llama.cpp and KTransformers support it locally, and Alibaba point to Qwen-Agent for tool calling.

Lineage

The model is the updated non-thinking version of Qwen3-30B-A3B. Alibaba list gains in instruction following, logical reasoning, text comprehension, mathematics, science, coding, tool use and long-tail multilingual knowledge, with better alignment on open-ended tasks.

Official Resources