Supported models

Every training run names two models. Use only the values on this page. A student_model_name or teacher_model_name from outside this catalogue either fails validation or is not available on the inference provider. It never falls back to a default.

• Student. The small model we fine-tune for your task. This is the one you deploy and run inference with, so pick a size that fits your latency, cost and hardware constraints.
• Teacher. The large model that generates and validates your synthetic training data. Its knowledge is distilled into the student, and you never deploy it.

Student models

Model	Value
Llama 3.2 1B Instruct	Llama-3.2-1B-Instruct
Llama 3.2 3B Instruct	Llama-3.2-3B-Instruct
Llama 3.1 8B Instruct	Llama-3.1-8B-Instruct
SmolLM2 135M	SmolLM2-135M-Instruct
SmolLM2 1.7B	SmolLM2-1.7B-Instruct
FunctionGemma 270M	functiongemma-270m-it
Gemma 3 270M	gemma-3-270m-it
Gemma 3 1B	gemma-3-1b-it
Gemma 3 4B	gemma-3-4b-it
Gemma 4 E2B	gemma-4-E2B-it
Gemma 4 E4B	gemma-4-E4B-it
Qwen3 0.6B	Qwen3-0.6B
Qwen3 1.7B	Qwen3-1.7B
Qwen3 4B	Qwen3-4B-Instruct-2507
Qwen3 8B	Qwen3-8B
Qwen3.5 0.8B	Qwen3.5-0.8B
Qwen3.5 2B	Qwen3.5-2B
Qwen3.5 4B	Qwen3.5-4B
Qwen3.5 9B	Qwen3.5-9B
Liquid LFM2 350M	LFM2-350M
Liquid LFM2 1.2B	LFM2-1.2B
Liquid LFM2 2.6B	LFM2-2.6B
Liquid LFM2.5 350M	LFM2.5-350M
Liquid LFM2.5 1.2B Instruct	LFM2.5-1.2B-Instruct
Qwen3.6 35B A3B	Qwen3.6-35B-A3B
Nemotron 3.5 Lightning 30B A3B	Nemotron-3.5-Lightning-30B-A3B

Start with Qwen3-1.7B, read the metrics, and only move up in size if the results fall short. Students under 1B parameters give the fastest and cheapest inference and are the easiest to run on-device, but need more training data to reach a given quality bar. Students of 3B and up usually reach higher quality with less work, at slower and more expensive inference.

Trying several is cheap. A sweep is one submission per student against the same training dataset, so it costs training credits only. See Training several students at once.

Teacher models

Every teacher below is an open-weight model.

Model	Value	Tool calling	Vision
GPT OSS 120B	openai.gpt-oss-120b	✓	✗
GPT OSS 120B Thinking	openai.gpt-oss-120b-thinking	✓	✗
GPT OSS 20B	openai.gpt-oss-20b	✓	✗
GPT OSS 20B Thinking	openai.gpt-oss-20b-thinking	✓	✗
DeepSeek V3.1	deepseek.v3.1	✗	✗
DeepSeek V4 Pro	deepseek.v4-pro	✓	✗
DeepSeek V4 Pro Thinking	deepseek.v4-pro-thinking	✓	✗
DeepSeek V4 Pro 0813 Low Thinking	deepseek.v4-pro-0813-low-thinking	✓	✗
DeepSeek V4 Pro 0813 High Thinking	deepseek.v4-pro-0813-high-thinking	✓	✗
DeepSeek V4.1 Flash Low Thinking	deepseek.v4.1-flash-low-thinking	✓	✓
DeepSeek V4.1 Flash High Thinking	deepseek.v4.1-flash-high-thinking	✓	✓
Qwen2.5 VL 72B	Qwen2.5-VL-72B-Instruct	✗	✓
Qwen3 235B A22B	Qwen3-235B-A22B-Instruct-2507	✓	✗
Qwen3 480B A35B Coder	Qwen3-480B-A35B-Coder	✗	✗
Qwen3.8 2.4T A95B Minimal Thinking	Qwen3.8-2.4T-A95B-minimal-thinking	✓	✗
Qwen3.8 2.4T A95B Medium Thinking	Qwen3.8-2.4T-A95B-medium-thinking	✓	✗
ZAI GLM 5	zai.glm-5	✓	✗
ZAI GLM 5 Thinking	zai.glm-5-thinking	✓	✗
ZAI GLM 5.2	zai.glm-5.2	✓	✗
ZAI GLM 5.2 Thinking	zai.glm-5.2-thinking	✓	✗
ZAI GLM 5.3 Low Thinking	zai.glm-5.3-low-thinking	✓	✗
ZAI GLM 5.3 High Thinking	zai.glm-5.3-high-thinking	✓	✗
ZAI GLM 5.3 Flash Low Thinking	zai.glm-5.3-flash-low-thinking	✓	✓
ZAI GLM 5.3 Flash High Thinking	zai.glm-5.3-flash-high-thinking	✓	✓
Moonshot Kimi K2.6	moonshotai.kimi-k2.6	✓	✓
Moonshot Kimi K2.6 Thinking	moonshotai.kimi-k2.6-thinking	✓	✓
Moonshot Kimi K3	moonshotai.kimi-k3	✓	✓
Moonshot Kimi K3 Low Thinking	moonshotai.kimi-k3-low-thinking	✓	✓
Moonshot Kimi K3 Max Thinking	moonshotai.kimi-k3-max-thinking	✓	✓
NVIDIA Nemotron 3 Ultra	nvidia.nemotron-3-ultra	✓	✗

The tool calling column covers tool-calling-closed-book, multi-turn-tool-calling-closed-book, and any chat completion task whose job description declares tools. The vision column covers base.visual_task.

Some families give a plain value and a -thinking value. The two point at the same model. The -thinking value turns the reasoning mode on, and the plain value turns it off. Newer families name the reasoning effort instead: -low-thinking and -high-thinking, -minimal-thinking and -medium-thinking, or -low-thinking and -max-thinking. A higher effort gives better answers on hard tasks, but it is slower and it costs more. moonshotai.kimi-k3-thinking is an old name for moonshotai.kimi-k3-max-thinking, and a config that still uses it runs with a warning.

The default teacher, openai.gpt-oss-120b, is a strong choice for every task type. The teacher is also the main lever when teacher evaluation comes back short, so this is the table to come back to.

Task compatibility

Not every model works with every task type.

• Tool-calling students. Only the Qwen3, Qwen3.5, Qwen3.6, Llama 3, LFM2, LFM2.5, FunctionGemma, Gemma 4 and Nemotron models handle tool-calling-closed-book, multi-turn-tool-calling-closed-book, and the chat completion tasks when tools are declared. A chat-completion job with no tools has no such restriction.
• Tool-calling teachers. Only the ones marked ✓ above.
• Vision teachers. base.visual_task: true needs a teacher marked ✓ in the vision column, and the same rule applies to trace_processing.teacher_model_name, evaluation.llm_as_a_judge_model_name and every model in trace_processing.relabelling_committee_models. One non-vision model in any of these roles fails validation.
• Vision students. base.visual_task: true also needs a Qwen3.5 or Gemma 4 student.
• Reasoning teachers. Every teacher except Qwen3-235B-A22B-Instruct-2507, Qwen3-480B-A35B-Coder and Qwen2.5-VL-72B-Instruct needs synthgen.teacher_temperature between 0.5 and 0.7. A value outside that range fails validation.
• GPT OSS reasoning effort. openai.gpt-oss-20b and openai.gpt-oss-120b run at low reasoning effort. Their -thinking values run at medium, so they reason for longer before they answer.

Using these in your config

Set both in the base section, using the exact values from the tables above:

base:
  task: question-answering
  student_model_name: Llama-3.2-1B-Instruct
  teacher_model_name: openai.gpt-oss-120b

Next

Every other parameter and its default: Config file.