Leading general-purpose foundation models & professional AI agents. The proprietary Spark model family, with AStudio & Loomy harnesses. Open source.
Joined August 2026
- Tweets115
- Following22
- Followers1.8K
- Likes161
Spark-X2.6-4B is here! Stronger coding, agents & long-context skills through improved post-training. Thinking: none / low / high.
@huggingface @ModelScope2022 @ollama @vllm_project @sgl_project
huggingface.co/XHToken/Spark…
Selected release results (high thinking):
BFCL-V4: 71.8
MCP-Atlas: 62.4
BrowseComp: 53.1
SWE Pro: 54.3
The attached full table preserves evaluation notes. The inference chart compares hardware, runtimes and context lengths.
Get Spark-X2.6:
Hugging Face: huggingface.co/collections/X…
ModelScope: modelscope.cn/collections/XH…
Ollama: ollama.com/SparkLLM/Spark-X2…
Run locally:
ollama run SparkLLM/Spark-X2.6-4B
Deployment guides include vLLM and SGLang.
To repeat: deploy from the project docs, open My Skills, then check the saved avatar, coordinate and version. Invitation history is in team member management. Release: github.com/iflytek/skillhub/…
A synthetic insert needs area 40 cm² and a length 3 cm greater than its width. Local Spark-X2.5-1.7B extracts the integers; SymPy solves. Demo: youtube.com/shorts/QYTkE-Np6…
Both −8 and 5 satisfy x(x+3)=40. Requiring positive width and length retains 5 × 8 cm. Windows, llama.cpp b10833, community Q4_K_M, Python 3.12 and SymPy 1.14.0. One actual model request.
Follow the public SymPy skill, solve over the reals, substitute each root, then check the original physical conditions. Local demonstration adapter. github.com/K-Dense-AI/scient… #SparkX25 #SymPy
A negative root can satisfy an equation and still be an invalid size. In our local synthetic packaging demo, Spark-X2.5-1.7B extracted area 40 and gap 3; SymPy returned −8 and 5. Public skill: github.com/K-Dense-AI/scient…
The local adapter validates integer JSON, constructs x(x+3)−40 from SymPy symbols and exact Integers, and solves over S.Reals. Both roots give residual 0; requiring positive width and length leaves 5 × 8 cm.
One actual model request: Windows, llama.cpp b10833, community Q4_K_M weights, SymPy 1.14.0. Reuse the skill, solve, substitute, and check physical constraints. huggingface.co/XHToken/Spark… #SparkX25 #SymPy
Correct intermediate arithmetic, corrupted final number: 0.538461538 becomes 538,461,538 after a decimal disappears. Watch dearankh’s Spark-X2.5-1.7B Q4_K_M case: youtube.com/shorts/c9NIbxcVF… #SparkX25 #LLM
The disclosed Codex-operated evaluation used CPU and llama.cpp, one sample per item, and a 96-token reasoning budget. Original scoring: 19/28 items pass; 8/14 pairs pass both. The video recounts public records.
A small diagnostic of one configuration: final numbers and units need checking even when intermediate arithmetic is right. Original submission: huggingface.co/XHToken/Spark… #ModelEvaluation
An empty list crashed our scoring test. A local OpenCodeReview + Spark-X2.5-1.7B scan mentioned the missing guard. Watch the demo: youtube.com/watch?v=u2yBKugF…
Recorded on Windows: OpenCodeReview v1.13.0, llama.cpp and community Q4_K_M GGUF. After we manually added the guard, [] returned 0.0 and [2,4] returned 3.0; both assertions passed.
To try the same checks, use ocr scan --preview to verify the files, then test the review comments against boundary inputs. Source: github.com/alibaba/open-code… #SparkX25 #CodeReview
First preview the target with ocr scan --repo <test-repo> --path score.py --preview, then check comments with boundary inputs. github.com/alibaba/open-code… #SparkX25 #CodeReview
Spark X2.5 just crossed 2M+ downloads 💜
1M → 2M in only 15 days. Thank you to every developer who downloaded, quantized and built with it 🙏
And the next drop is almost here 🎁 Spark-X2.6-4B / 1.7B goes open source soon.
What should we upgrade? Tell us 👇
Spark-X2.6-4B / 1.7B will soon be open source. We’re excited to explore what’s possible with local AI, together with you. @SparkLLM
See you next week!
2,029,840 downloads, by the numbers:
📦 1,019,681 across 6 official channels: Hugging Face, Ollama, ModelScope, Modelers, AtomGit, SCNet
🛠️ 1,010,159 from 109 community builds: GGUF, MLX, ONNX and more
Try the 1.7B & 4B models 👇
huggingface.co/collections/X…
