The second model in the SparkLLM team's Spark-X2.6 series. It shares the same architecture as Spark-X2.5-1.7B, with enhanced programming, agentic, long-context, and multilingual capabilities achieved through improved post-training. The 4B version of the same series was included on 10 24; this 1.7B version only had its weights uploaded on 10 27, so it was not included in previous editions.
Score increase 56 · Likes 58 · Downloads 1307
Reasons
- Smaller than the 4B version. The authors claim that on an RTX 4090, decoding speed for short contexts leads Qwen3.5-2B, MiniCPM-2B, and LFM2.5-2.6B, and with their optimized llama.cpp fork, it leads across all tested lengths (charts only).
- Thinking budget offers three options: none / low / high. Additionally, there is a community-made WebGPU demo (mohdmahmodi/Spark-X2.6-1.7B-WebGPU) that runs locally in the browser using the official INT4 weights, requiring no installation.
- Comprehensive deployment options: Documentation is available for Transformers, SGLang, vLLM (requires installing the Spark-plugin), MLX, llama.cpp, Ollama, and LM Studio. It also supports Ascend NPUs.
Usage
- ollama run SparkLLM/Spark-X2.6-1.7B (requires Ollama 0.34.1 or higher); you can also search for Spark-X2.6 in LM Studio 0.4.0 or higher; llama.cpp requires b10828 or higher.
- Transformers: Install 4.57.1, enable trust_remote_code=True, and use the enable_thinking switch in apply_chat_template to toggle thinking; recommended temperature 1.0, top_p 0.95.
- SGLang: Add --tool-call-parser spark25 --reasoning-parser qwen3; to disable thinking, set enable_thinking=false in the request's chat_template_kwargs.
Risks
- Benchmark results and speed comparisons are provided only as images; the model card text does not contain any specific scores, so this column cannot verify claims of "leading within its size class."
- The SGLang example sets the context length to 1048576; if VRAM is insufficient, reduce --context-length; the authors did not provide specific memory usage details.
- The WebGPU demo is a community project, tested only in the contributor's own Windows + NVIDIA environment.