Qwen3 30B A3B Instruct 2507

Qwen3 30B A3B Instruct 2507 is a causal language model released in July 2025 as the updated non-thinking variant in the Qwen3 family, refined through both pretraining and post-training stages. It is built on a mixture-of-experts architecture totaling 30.5 billion parameters, with only 3.3 billion activated at inference and 29.9 billion non-embedding parameters. The structure spans 48 layers with grouped-query attention configured at 32 query heads and 4 key-value heads, drawing on a pool of 128 experts of which 8 are activated per token, a design that aims to balance computational efficiency with broad capability coverage.

The model targets practical, instruction-driven use cases, with documented gains in instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage, along with improved long-tail knowledge across multiple languages and better alignment on subjective, open-ended prompts. It also supports natively extended context understanding suitable for long-document workflows. Because it operates exclusively in non-thinking mode and does not emit think blocks in outputs, it is well suited for deployments that want direct, immediately usable responses rather than chain-of-thought traces, and it integrates with OpenRouter API access and Hugging Face hosting for flexible serving.

CoreWeaveQwen/Qwen3-30B-A3B-Instruct-2507qwen

Open provider page Provider docs

Quick Info

Provider: CoreWeave
Model key: Qwen/Qwen3-30B-A3B-Instruct-2507
Release date: Jul 29, 2025
Last updated: Jul 29, 2025
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.10
Output token cost: $0.30

Limits

Output tokens: 262,144 tokens
Context window: 262,144 tokens

Transparent token rates

Compare Qwen3 30B A3B Instruct 2507 pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Qwen3 30B A3B Instruct 2507

Weights & Biases

CoverageDiscourse

This is the official HuggingFace discussions index for the canonical Qwen/Qwen3-30B-A3B-Instruct-2507 repository, confirming the exact variant exists at that path and has an active community. Open and closed threads cover topics such as recommended GPU setups for Qwen3 30B, an external OMS scoring post (OMS 70.4 B), GP Earlier threads also include user questions on static KV-cache support, Polish language coverage, LoRA fine-tuning issues, and installation guides. While the page itself is a discussion index rather than a substantive announcement, it independently corroborates the model's identity and surfaces ecosystem activity aroun

Weights & Biases

Coverage

This Ollama page hosts a third-party community mirror (alibayram/Qwen3-30B-A3B-Instruct-2507) of the exact Qwen variant, published about a year ago as a Q4_K_M quantization. The reproduced README describes Qwen3-30B-A3B-Instruct-2507 as a causal language model at the post-training stage with 30.5B total parameters and According to the same listing, Qwen frames this release as an updated non-thinking variant of the earlier Qwen3-30B-A3B, with stated gains in instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool usage, broader multilingual long-tail knowledge, better subjective/open-ended

Weights & Biases

Coverage

The Hugging Face model card for Qwen/Qwen3-30B-A3B-Instruct-2507 explicitly names the exact subject variant and serves as the strongest first-party technical reference among the supplied candidates. It documents the architecture as a causal language model with 30.5B total parameters and 3.3B activated parameters across The card lists enhancement highlights versus the prior Qwen3-30B-A3B release: significant gains in instruction following, logical reasoning, mathematics, science, coding, and tool usage; broader long-tail multilingual knowledge; better subjective/open-ended alignment; and enhanced 256K long-context understanding. It al