We collect reasoning, coding, math, and agentic traces from frontier models and package them as open datasets for training the next generation of efficient models.
Distillation traces from 8 frontier models. Reasoning, coding, math, agentic — all in one dataset.
Traces from GLM 5.2, Fable 5, GPT-5.6 Sol, DeepSeek V4 Pro, Qwen 3.8 Max, Kimi K3, and Manus. Parquet format. Built for SFT distillation training. Created August 21, 2026.
Agentic and tool-use traces from Opus 4.7, Kimi 2.6, GLM 5.2, DeepSeek 4, Mythos 5, Fable 5, Qwen 3.7. 294 downloads.
2M+ reasoning traces from FABLE 5. 28 downloads. The raw material for distillation.
MLX and GGUF quantized builds — optimized for local inference.
Traces across reasoning, coding, math, and agentic tasks from 8 source models.
We run inference across 8+ frontier models — GLM, GPT, DeepSeek, Qwen, Kimi, Manus — each bringing different reasoning strengths.
Reasoning chains, code outputs, math solutions, agentic tool calls — structured for SFT training in parquet format.
Every dataset published on Hugging Face. Use them to train your own models — reasoning, coding, math, or agentic.
Solstice-AI collects reasoning, coding, math, and agentic traces from frontier models and packages them as open datasets for SFT training.
Our Solace 1.0 dataset combines traces from 8 different frontier models — each contributing unique reasoning patterns. Our Axiom dataset focuses on agentic and tool-use traces.
Every dataset is published openly. We believe the raw material for training better models should be accessible to everyone.
Latest distillation traces refreshed — 314 downloads and counting.
Traces from 8 frontier models. Reasoning, coding, math, agentic. AGPL-3.0.
MLX 6-bit quantized build — 302 downloads. Apple Silicon optimized.
Building the future of efficient AI.
Build pipelines for collecting and curating frontier model traces.
Scale trace ingestion, formatting, and quality filtering.
Early-stage. Whatever's needed.