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 catalog fails validation.

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

Recommended models

Student: the Qwen3.5 family. Start with Qwen3.5-4B. Go smaller (Qwen3.5-2B, Qwen3.5-0.8B) when your deployment target needs it, or larger (Qwen3.5-9B) when accuracy matters more than cost.

Teachers: GLM 5.3. zai.glm-5.3-flash-low-thinking is the small teacher, for simple tasks and for cost. zai.glm-5.3-low-thinking is the large teacher, for everything else. Move either to its -high-thinking value when teacher evaluation shows the task needs more reasoning.

The judge. evaluation.llm_as_a_judge_model_name scores every run, and scores from two judges aren’t comparable, so it is usually set in the first config and kept (Variation between runs). When it’s omitted, it takes the value of base.teacher_model_name at the time the job is created and is written into the config, so a config read back for an override names it, and changing the teacher there leaves the judge as it is. trace_processing.teacher_model_name, the relabeling model, works the same way.

When the fields are omitted, the config defaults are Llama-3.2-1B-Instruct and openai.gpt-oss-120b. A config with the recommended models:

base:
  task: question-answering
  student_model_name: Qwen3.5-4B
  teacher_model_name: zai.glm-5.3-low-thinking
evaluation:
  llm_as_a_judge_model_name: zai.glm-5.3-low-thinking

Student models

Family	Values	Tool calling
Qwen3.5	Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B, Qwen3.5-9B	✓
Llama 3	Llama-3.2-1B-Instruct, Llama-3.2-3B-Instruct, Llama-3.1-8B-Instruct	✓
Qwen3	Qwen3-0.6B, Qwen3-1.7B, Qwen3-4B-Instruct-2507, Qwen3-8B	✓
LFM2 / LFM2.5	LFM2-350M, LFM2-1.2B, LFM2-2.6B, LFM2.5-350M, LFM2.5-1.2B-Instruct	✓
Gemma 4	gemma-4-E2B-it, gemma-4-E4B-it	✓
Qwen3.6	Qwen3.6-35B-A3B	✓
Qwen3.8	Qwen3.8-27B (QLoRA only)	✓
FunctionGemma	functiongemma-270m-it	✓
Gemma 3	gemma-3-270m-it, gemma-3-1b-it, gemma-3-4b-it	✗
SmolLM2	SmolLM2-135M-Instruct, SmolLM2-1.7B-Instruct	✗

Qwen3.8-27B can only be trained with tuning.use_qlora: true. Config validation doesn’t check this, so set it in every config that trains it.

Size tiers

Size range	When to choose
135M - 350M	Extreme latency constraints, on-device, very simple tasks
0.6B - 1.2B	Cost-sensitive production, well-defined tasks
1.7B - 3B	Better quality than the 1B class within a small memory and latency budget
4B	The default: start here
8B - 9B	Accuracy-critical tasks, multi-step reasoning, nuanced outputs
27B dense	Highest quality from a dense model; needs a large GPU to serve
35B MoE (3B active)	Highest quality; inference costs as much as a 3B model, but the weights need the memory of a 35B one

Trying several sizes means one training run per student against the same training dataset. See Model training.

Teacher models

Every teacher is an open-weight model.

Family	Values
GLM	zai.glm-5.3-low-thinking, zai.glm-5.3-high-thinking, zai.glm-5.3-flash-low-thinking, zai.glm-5.3-flash-high-thinking
GPT OSS	openai.gpt-oss-20b, openai.gpt-oss-20b-thinking, openai.gpt-oss-120b, openai.gpt-oss-120b-thinking
DeepSeek	deepseek.v3.1, deepseek.v4-pro, deepseek.v4-pro-thinking, deepseek.v4-pro-0813-low-thinking, deepseek.v4-pro-0813-high-thinking, deepseek.v4.1-flash-low-thinking, deepseek.v4.1-flash-high-thinking
Qwen	Qwen3-235B-A22B-Instruct-2507, Qwen3-480B-A35B-Coder, Qwen2.5-VL-72B-Instruct, Qwen3.8-2.4T-A95B-minimal-thinking, Qwen3.8-2.4T-A95B-medium-thinking
Kimi	moonshotai.kimi-k2.6, moonshotai.kimi-k2.6-thinking, moonshotai.kimi-k3, moonshotai.kimi-k3-low-thinking, moonshotai.kimi-k3-max-thinking

A plain value and its -thinking value point at the same model: the -thinking value turns the reasoning mode on, the plain value turns it off. Some families name the reasoning effort instead: -low-thinking and -high-thinking (GLM 5.3, DeepSeek V4 Pro 0813, DeepSeek V4.1 Flash), -minimal-thinking and -medium-thinking (Qwen3.8), -low-thinking and -max-thinking (Kimi K3). A higher effort gives better answers on hard tasks, and is slower and costs more.

openai.gpt-oss-20b and openai.gpt-oss-120b run at low reasoning effort; their -thinking values run at medium.

Every teacher is a reasoning model except Qwen3-235B-A22B-Instruct-2507, Qwen3-480B-A35B-Coder and Qwen2.5-VL-72B-Instruct. A reasoning teacher needs synthgen.teacher_temperature between 0.5 and 0.7 (Config file).

Reasoning students

The students that support base.enable_thinking: Reasoning models.

Vision

base.visual_task: true needs a vision-capable model in every teacher and judge role: base.teacher_model_name, 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 role fails validation.

• Vision teachers: moonshotai.kimi-k2.6, moonshotai.kimi-k2.6-thinking, moonshotai.kimi-k3, moonshotai.kimi-k3-low-thinking, moonshotai.kimi-k3-max-thinking, zai.glm-5.3-flash-low-thinking, zai.glm-5.3-flash-high-thinking, deepseek.v4.1-flash-low-thinking, deepseek.v4.1-flash-high-thinking, Qwen2.5-VL-72B-Instruct.
• Vision students: Qwen3.5-0.8B, Qwen3.5-2B, Qwen3.5-4B, Qwen3.5-9B, Qwen3.8-27B, gemma-4-E2B-it, gemma-4-E4B-it.

Tool-calling compatibility

This applies to the chat completion tasks whenever the job description declares tools (always, for chat-completion-agentic). A chat-completion job with no tools has no model restriction.

• Students: only the families marked ✓ in the student table: Qwen3.5, Llama 3, Qwen3, LFM2/LFM2.5, Gemma 4, Qwen3.6, Qwen3.8 and FunctionGemma.
• Teachers: every teacher except deepseek.v3.1, Qwen3-480B-A35B-Coder and Qwen2.5-VL-72B-Instruct.

Next

Every other parameter and its default: Config file.