Qwen3 4B Instruct 2507
Qwen/Qwen3-4B-Instruct-2507
published Aug 2025 · updated Sep 2025
Qwen3 4B Instruct 2507 is a text-generation model that is an updated version of Qwen3-4B non-thinking mode, with significant improvements in instruction following, logical reasoning, coding, tool usage, and multilingual support.
specs
| Task | text-generation |
| Architecture | Causal Language Model with GQA (32 Q heads, 8 KV heads) |
| Parameters | 4.0B total (3.6B non-embedding) |
| Context Length | 262,144 tokens natively |
| License | Apache 2.0 |
about this model
Key Capabilities and Improvements
Substantial improvements have been made in long-tail knowledge coverage across 119 languages and dialects. Alignment with user preferences in subjective and open-ended tasks is markedly better, enabling more helpful responses and higher-quality text generation. The model also demonstrates enhanced 256K long-context understanding.
Benchmark Performance
The model achieves strong results across knowledge, reasoning, coding, alignment, and agentic tasks. Selected scores include:
- Knowledge: MMLU-Pro 69.6, GPQA 62.0
- Reasoning: AIME25 47.4, LiveBench (20241125) 63.0
- Coding: LiveCodeBench v6 35.1, MultiPL-E 76.8
- Alignment: Arena-Hard v2 43.4, Creative Writing v3 83.5
- Agent: BFCL-v3 61.9, TAU1-Retail 48.7
- Multilingual: MultiIF 69.0, PolyMATH 31.1
This model is hosted by gigarouter as an OpenAI-compatible API, providing convenient access for text-generation workloads without the need for manual deployment.
best for
- ·Instruction following and open-ended text generation
- ·Logical reasoning and math problem solving
- ·Coding and tool usage with function calling
- ·Multilingual long-context understanding
FAQ
It excels at instruction following, logical reasoning, coding, tool usage, and multilingual tasks, with improved alignment and higher-quality text generation.
This updated version shows significant gains across benchmarks, including MMLU-Pro (58.0 to 69.6), AIME25 (19.1 to 47.4), and Arena-Hard v2 (9.5 to 43.4).
It is released under the Apache 2.0 license.
It uses the standard chat template with user/assistant messages. The model supports non-thinking mode only, so no thinking blocks are generated.
Use the gigarouter OpenAI-compatible endpoint with your API key. The endpoint supports chat completions; pass the model name as qwen3-4b-instruct-2507.
# rerank documents by relevance; billed per document curl https://gigarouter.ai/v1/rerank \ -H "Authorization: Bearer $GR_KEY" \ -d '{"model":"Qwen/Qwen3-4B-Instruct-2507","query":"capital of France", "documents":["Paris is the capital of France.","Bananas are yellow."]}'
try it live
runs the real hosted model on a shared demo allowance · get your own key + $25 free →