Spark-X2.6-1.7B is here: a compact model for on-device AI, following Spark-X2.6-4B. If you find it useful, please give the model a Like on Hugging Face and share your feedback.
What's new
The model keeps the architecture of Spark-X2.5-1.7B and improves its capabilities through post-training. The official model card describes gains in coding, agent tasks, long-context reasoning, multilingual tasks and reasoning, with three configurable thinking budgets: none, low and high. The model is licensed under Apache 2.0.
Selected results
The official benchmark table reports these results for X2.5-1.7B → X2.6-1.7B:
| Benchmark | X2.5-1.7B | X2.6-1.7B |
|---|---|---|
| SWE Verified | 28.3 | 59.8 |
| SWE Multilingual | 23.3 | 57.3 |
| τ²-bench | 65.3 | 85.3 |
These are model-team evaluations. The table uses thinking mode, with X2.6 evaluated at high. Asterisks in the full chart identify externally reported scores; SWebBench and PAgentBench are in-house benchmarks. See the retained notes for sampling settings and evaluation details.
Local inference
The second figure separates prefill and decode throughput at 4K, 32K and 128K context lengths. Its Mac M5 Pro 16-core group uses MLX INT4 weights and BF16 KV cache. Its RTX 4090 group uses llama.cpp Q4_K_M weights and FP16 KV cache, with an additional Spark variant using the team's optimized llama.cpp implementation. The software versions and quantization details are in the figure. Results depend on the tested configuration.
Choose your model and runtime
Hugging Face lists the post-trained model, Base, GGUF, INT8, INT4 and FP8 repositories. The supplied ModelScope screenshot shows the post-trained, Base, INT8, INT4 and FP8 options. Choose the format supported by your runtime.
- Model page — leave a Like here
- Hugging Face collection
- ModelScope collection
- Spark-X2.6-1.7B on Ollama
Run with Ollama:
ollama run SparkLLM/Spark-X2.6-1.7B
Four launch images
In order: benchmarks, inference throughput, Hugging Face model list, ModelScope model list. The platform screenshots are launch snapshots; counters and relative timestamps can change.
Try it on a real task, tell us your hardware and runtime, and share what you build. Thank you for supporting open models!
Top comments (0)