Alibaba's Qwen3-4B-Instruct-2507 Jumps 45 Points on Logic, Fits in 3GB
Unsloth's dynamic GGUF quants let you run Qwen's updated 4B instruct model locally with 256K context on modest hardware.
- Unsloth released dynamic GGUF quants of Qwen3-4B-Instruct-2507, now at 838K downloads.
- MMLU-Pro climbs to 69.6, beating GPT-4.1-nano's 62.8 at 4B parameters.
- AIME25 math score jumps from 19.1 to 47.4 versus the prior Qwen3-4B.
- Native 256K context window for long-document RAG and summarization workloads.
- UD-Q4_K_XL quant runs on 8GB RAM via llama.cpp, Ollama, or Docker Model Runner.
- Weak on agentic benchmarks: TerminalBench Hard 4.5%, TAU2 26.6%, Aider-Polyglot 12.9.
Alibaba’s Qwen team released Qwen3-4B-Instruct-2507 in July 2025, and Unsloth converted the checkpoint into GGUF files for llama.cpp, Ollama, and Docker Model Runner. Unsloth’s GGUF repository displayed more than 838,000 downloads at the time of writing. Selected quantizations occupy roughly 2 to 3 GB, giving developers a practical local model with stronger reasoning, coding, and instruction-following scores than the previous 4B release.
The July checkpoint lifts 4B performance
Qwen3-4B-Instruct-2507 is a dense, 4 billion-parameter model tuned for direct responses without Qwen’s optional thinking mode. The Qwen model card lists a 262,144-token context window, multilingual support, tool-use capabilities, and an Apache 2.0 license.
| Property | Value |
|---|---|
| Parameters | 4 billion |
| Architecture | Dense, decoder-only transformer |
| Response mode | Instruct, non-thinking |
| Maximum context | 262,144 tokens |
| License | Apache 2.0 |
Qwen’s published evaluations show substantial gains over the previous Qwen3-4B checkpoint running in non-thinking mode. The scores below use Qwen’s test settings and represent percentage-point changes where applicable.
| Benchmark | Previous 4B | 2507 checkpoint | Change |
|---|---|---|---|
| MMLU-Pro | 58.0 | 69.6 | +11.6 |
| GPQA | 41.7 | 62.0 | +20.3 |
| AIME25 | 19.1 | 47.4 | +28.3 |
| ZebraLogic | 35.2 | 80.2 | +45.0 |
| LiveCodeBench v6 | 26.4 | 35.1 | +8.7 |
The largest gains appear in logic, mathematics, and scientific reasoning. Qwen also reports an MMLU-Pro score of 62.8 for GPT-4.1-nano under its comparison setup, placing the refreshed 4B checkpoint 6.8 points higher on that test. Application-specific evaluation remains necessary because prompts, quantization levels, sampling settings, and retrieval quality can shift production results.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
