SparkLLM 發佈 Spark-X2.6-4B，強化端側智能體與編碼能力
Spark-X2.6-4B: stronger on-device agents, coding and controllable thinking
SparkLLM 發佈 Spark-X2.6-4B，並表示後訓練提升了編碼、Agent 工作流程、長上下文、多語言使用及推理能力。在思考等級設為 high 的公佈評測中，BFCL-V4 由前代的 65.1 升至 71.8，MCP-Atlas 由 54.6 升至 62.4，SWE Verified 由 41.6 升至 62.8；各基準排名不一，數據由發布團隊提供。
Spark-X2.6-4B is here: our latest compact, general-purpose language model for on-device agents and real-world tasks.
It keeps the architecture of Spark-X2.5-4B and improves capabilities through post-training, with gains in coding, agent workflows, long-context tasks, multilingual use and reasoning.
Stronger agents and coding in 4B parameters
The release focuses on tool use, multi-step planning and software engineering. Selected results from our published evaluation:
| Benchmark | Spark-X2.6-4B | Spark-X2.5-4B |
|---|---|---|
| BFCL-V4 | 71.8 | 65.1 |
| MCP-Atlas | 62.4 | 54.6 |
| BrowseComp | 53.1 | 40.9 |
| SWE Pro | 54.3 | 44.4 |
| SWE Verified | 62.8 | 41.6 |
| AA-LCRv1.1 | 67.3 | 64.7 |
The full comparison includes similarly sized models and larger 9B and 12B models. Spark-X2.6-4B leads on several listed tasks; rankings vary by benchmark.
These are the release team's reported results. The table evaluates Spark-X2.6-4B with thinking set to high. Asterisks identify comparison scores taken from public model cards or papers; dashes mean scores are unavailable. The original footnotes are preserved in the image.
A controllable thinking budget
Three thinking levels—none, low and high—let applications balance reasoning effort, latency and inference cost. Use the model card's configuration guidance for the selected runtime.
On-device inference
The inference chart reports prefill and decode throughput on RTX 4090 and Mac M5 Pro (16-core) across context lengths. It labels each runtime and quantization setting, including llama.cpp Q4_K_M and Spark INT4 runs using TensorRT or MLX. Compare the relevant hardware, runtime and context length when choosing a deployment.
Download and run
| Platform | Download | Description |
|---|---|---|
| Hugging Face | Spark-X2.6 in Huggingface | Official Hugging Face model collection for Spark-X2.6 |
| ModelScope | Spark-X2.6 in ModelScope | Recommended download source for users in China |
| Ollama | Spark-X2.6 in Ollama | Download and run the model locally with Ollama |
Direct model weights and model card: https://huggingface.co/XHToken/Spark-X2.6-4B.
With Ollama installed, run:
ollama run SparkLLM/Spark-X2.6-4B
The release documentation also covers Transformers, SGLang, vLLM, MLX and llama.cpp. Follow the runtime-specific setup in the model card for required versions, images, plugins and hardware configuration.
The model is released under Apache 2.0. Try it on a coding task or tool-use workflow and share your runtime, hardware and results with the SparkLLM community.
來源：DEV Community · dev.to