SparkLLM's Spark-X2.6 Doubles Coding Scores in a Tiny 1.7B On-Device Model
SparkLLM's new 1.7B on-device model nearly matches its 4B sibling on coding and agent benchmarks, with native 1M context and controllable thinking.
Takeaways
- SparkLLM released Spark-X2.6-1.7B, an Apache-2.0 on-device model with 1M native context.
- Hits SWE Verified 59.8 and tau2-bench 85.3, close to the 4B sibling.
- Same architecture as X2.5-1.7B; gains come entirely from improved post-training.
- Controllable thinking budget with none, low, and high levels to trade quality for latency.
- Day-0 support in vLLM and SGLang, plus Ollama, MLX and llama.cpp for local use.
- Ships in post-trained, Base, GGUF, INT8, INT4 and FP8 variants on HF and ModelScope.
SparkLLM has released Spark-X2.6-1.7B, a compact open-weight model aimed at on-device coding, tool use, and long-context applications. In SparkLLM’s high-thinking evaluations, the 1.7B model trails its 4B sibling by 3 points on SWE-bench Verified and 1.6 points on tau2-bench. Those narrow gaps make the smaller checkpoint relevant to developers constrained by memory, compute, or distribution size.
The Apache 2.0 release includes post-trained, base, GGUF, INT8, INT4, and FP8 variants across Hugging Face and ModelScope.
| Parameters | 1.7 billion |
|---|---|
| Native context | 1 million tokens |
| Thinking levels | None, low, and high |
| Primary languages | English and Chinese |
| License | Apache 2.0 |
Post-training drives the gains
SparkLLM attributes the capability increase to an improved post-training recipe. X2.6 retains the architecture of Spark-X2.5-1.7B while raising the team’s reported scores across coding, tool use, multilingual tasks, and reasoning.
| Benchmark | X2.6-1.7B | X2.6-4B | Gap |
|---|---|---|---|
| SWE-bench Verified | 59.8 | 62.8 | 3.0 |
| tau2-bench | 85.3 | 86.9 | 1.6 |
The same-size generational comparison shows how much the post-training update contributes, although the release specifies high thinking only for X2.6.