Open distillation lab
Frontier reasoning, distilled into models you can run at home.
We capture reasoning and agent traces from 37 frontier models across 12 labs, fine-tune open bases like Qwen and Gemma on them, and publish everything: the weights, the GGUF quants, and the datasets themselves.
- Apache-2.0 weights
- GGUF for llama.cpp & Ollama
- Trained with Unsloth
- Traces generated with teich
All-time downloads
2.2M
85,178 in the last 30 days
Model releases
66
121 repositories
Datasets
54
reasoning, agent & chat traces
Frontier teachers
37
from 12 labs
GGUF quants shipped
457
59 releases
Followers on HF
1.5K
Likes
3.7K
Distilled from
37 frontier teachers
Fresh off the GPU
Latest releases
Every fine-tune ships as full weights plus GGUF quants. Click a card for quant lists, run commands, and the datasets it was trained on.
Anthropic · Claude Fable 5
Qwen3.8 27B
Anthropic · Claude Opus 4.6
MiniMax M2.7
Anthropic · Claude Fable 5
Gemma 4 31B
Anthropic · Claude Fable 5
Qwen3.5 9Bv1
Anthropic · Claude Fable 5
Qwen3.6 27B
Anthropic · Claude Opus
Qwen3.6 27Bv2
The distillation matrix
Every base we have trained, against every lab we have learned from.
Rows are open-weight base families, columns are the labs whose frontier models taught them. Hover a cell to see the releases, click to open them in the catalog.
| Base ↓ · Teacher lab → | Anthropic | OpenAI | DeepSeek | Moonshot | Stealth | MiniMax | Z.ai | Cohere | Mistral | xAI | Xiaomi | Total | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3 | 9 | 11 | 8 | 4 | 3 | 2 | 2 | 1 | 1 | 1 | 1 | 43 | |
| Gemma 4 | 5 | 5 | |||||||||||
| gpt-oss | 1 | 1 | 1 | 3 | |||||||||
| Qwen3.5 | 3 | 3 | |||||||||||
| Qwen3.6 | 3 | 3 | |||||||||||
| Nemotron Cascade | 2 | 2 | |||||||||||
| Devstral Small | 1 | 1 | |||||||||||
| GLM 4.7 Flash | 1 | 1 | |||||||||||
| LFM2.5 | 1 | 1 | |||||||||||
| MiniMax M2.7 | 1 | 1 | |||||||||||
| Mistral 7B | 1 | 1 | |||||||||||
| Nemotron Orchestrator | 1 | 1 | |||||||||||
| Qwen3.8 | 1 | 1 | |||||||||||
| Total | 26 | 11 | 9 | 6 | 3 | 3 | 2 | 2 | 1 | 1 | 1 | 1 | 66 |
Community favorites
What people actually run
Ranked by all-time downloads across Hugging Face, counted across every repository of a release.
Most downloaded models
all timeMost downloaded datasets
all time- 01claude-4.5-opus-high-reasoning-250x26.7K
- 02DeepSeek v4 Pro Agent Traces25.1K
- 03Claude-Opus-4.6-Reasoning-887x16.1K
- 04gemini-3-pro-preview-high-reasoning-1000x5.1K
- 05Claude-Sonnet-4.6-Reasoning-1100x4.8K
- 06Ox-Alpha Pi Agent Coding Traces4.2K
Likes and 30-day trends are in the catalog. Sort by “Trending” to see what people are pulling this week.
How a distill is made
From frontier trace to local GGUF in four steps.
The whole pipeline is open. The datasets are on Hugging Face, the tooling is on PyPI, and the recipes are in the model cards.
- 01
Curate prompts
Coding, math, science and open-ended tasks. For agent traces we point real repositories at the teacher and let it work.
prompts.jsonl
- 02
Generate traces
Frontier models answer with maximum reasoning effort. Thinking, tool calls and results are captured with teich and published as a dataset.
teich generate
- 03
Fine-tune an open base
Qwen, Gemma, GLM, gpt-oss and friends are trained with Unsloth. Response-only masking keeps the loss on what the teacher actually said.
mask_data()
- 04
Quantize and ship
Safetensors plus imatrix GGUF from Q2 to Q8, vision projectors when the base sees images, and an Ollama Modelfile. Apache-2.0.
llama-server -hf
Training data, published
Newest datasets
Reasoning traces, agent sessions and chat data from frontier models, released so you can train your own.
Claude Fable 5
Claude Fable 5 Cursor Traces
Fable-5-Cursor-Traces
244 Fable 5 Cursor agent sessions for training & research. This dataset has 244 Cursor sessions with Fable 5 at High/xHigh/Max effort levels for distillation. [!IMPORTANT] This dataset is compatible with Teich! Use it directly in your…
Ox Alpha
Ox-Alpha Pi Agent Coding Traces
Ox-Alpha-Pi-Traces
This directory contains raw agent trace files generated by teich. JSONL files: 2247 Model metadata: stealth/ox-alpha Domains and prompt distribution Topic Traces Games & simulation (headless) 196 Frontend & Node-testable web 159 Health…
Ox Alpha
Ox Alpha Chat Traces
Ox-Alpha-10k
10,005 single-turn prompts for text-response teacher generation Each row carries id, category, subcategory All data was gathered using stealth/ox-alpha via OpenRouter (reasoning effort high) Topic distribution Category Rows Share Coding…
Open-source tooling
teich — the toolkit behind every dataset here.
Generate agent and chat traces, extract sessions from Claude Code, Codex, Cursor, Pi and Hermes, normalize them into training-ready rows, and mask the loss so you only train on what the teacher actually said. Same pipeline we use, one pip install away.
- PyPI
- v0.3.6
- GitHub stars
- 148
- Datasets made
- 4+
$ pip install teich$ teich init my-project && cd my-project$ teich generate -c config.yaml✓ 4,006 traces → output/ tools.json · dataset card · sandbox snapshots$ teich extract claude --model fable-5✓ anonymized 244 sessions → data/ upload to Hugging Face? (y/N)$
Built in public
Come hang out, request a distill, or keep the lights on.
Discord
Distill requests, release announcements, training chatter. The fastest way to reach us.
Join the serverHugging Face
1.5K followers get every drop first. Weights, quants and datasets live here.
Follow TeichAISupport the lab
Every coffee turns into API credits and GPU hours.
We are a small team paying for frontier-model generation and training runs out of pocket. A single high-reasoning dataset can cost more than a month of rent in tokens. If our models save you time, consider chipping in, or drop a distill request in the Discord and tell us what to build next.