Skip to main content
Consider these factors when choosing a base model:
  • Training mode: Every model supports low-rank adapter (LoRA) training, but only some support full-weight training. Check the Training modes column in the catalog, or list models programmatically.
  • Context length: The trainer’s maximum sequence length varies by model, and differs between LoRA and full-weight training.
  • GPU type: Post-training currently runs only on B200 GPUs (B200-SXM). Support for B300 and GB300 GPUs is coming soon.
  • Adapter capacity: If you plan to train several adapters together, check how many LoRA sessions one resource accepts at once. The limit varies by model.
  • Deployment: Not every model’s checkpoints can be deployed on Together AI. If you plan to serve the trained model, check its row in Checkpoints and deployment.
Start with Qwen/Qwen3.8-27B to get your training loop working, then scale up.
Post-training currently trains and samples on text only, even on base models that read images, like Nemotron 3.5 Super VL. To fine-tune on images, use vision fine-tuning.

Catalog

This table lists the public B200 configurations and syncs daily from /v1/rl/supported-models. For the live catalog, list the models programmatically. How to read the columns:
  • Training modes: Models that list both LoRA and full support full-weight training. For LoRA-only models, use lora_enabled=True.
  • Max context: The longest sequence (prompt plus response) that one forward_backward sample can hold.
  • Generator context: The longest sequence (prompt plus completion) that a sample call can handle. Where it’s larger than Max context (LoRA), keep GRPO rollouts within the trainer’s limit, because forward_backward rejects longer samples.
  • LoRA sessions per resource: The number of adapter sessions one resource accepts at once, which caps how many adapters you can train together. The limit varies by model and GPU type, and the API doesn’t return it yet.

Single-session variants

If you run only one training session on a resource, use a single-session variant for better throughput. Each variant has the same training modes and context limits as its default model, but a resource created from it accepts only one session at a time. The following models have a variant:
  • Qwen3.6 35B A3B: Qwen/Qwen3.6-35B-A3B-single-session.
  • Qwen3.8 27B: Qwen/Qwen3.8-27B-single-session.

Model capabilities

Thinking behavior and deployment options differ by model. Check your model’s row before you set thinking options or plan a deployment.

Thinking and chat templates

The cookbook renders each conversation through the model’s codec and chat template, which decide which thinking settings a run accepts: How to read the columns:
  • Thinking: The template’s default behavior:
    • On: The model thinks by default, and you can turn thinking off.
    • Always: The template opens a reasoning block on every turn.
    • Adaptive only: The model decides on each turn whether to think, and the template has no other mode.
    • No switch: The template has no setting that turns thinking off.
  • thinking_effort: The reasoning budgets the template accepts. Setting it on a model marked Not supported, or to a level the model doesn’t list, raises ValueError.
  • Earlier reasoning: Whether the template keeps the reasoning of earlier assistant turns when it renders the history and preserve_thinking is unset. Stripping it splits each conversation into one training sequence per turn.
  • Codec: The class in post_training_cookbook.codecs that renders prompts and parses responses for the model’s family.
Other model-specific behavior:
  • Qwen models and Nemotron 3.5 Super VL: For data without reasoning, set enable_thinking=False. Otherwise, the template renders each prompt differently from its full conversation, so the conversation doesn’t train.
  • GLM models: For data without reasoning, leave enable_thinking unset. The codec renders that data with thinking on. Each turn ends with the next message’s role marker, so an assistant turn followed by another assistant turn has no stop token. Merge or split consecutive assistant turns before training. The final reply has no stop token either, so the cookbook appends <|endoftext|> to it. A loop that builds samples by hand must append it too.
  • Kimi K3.1 Pro: Its template also accepts thinking_effort="medium", but the cookbook’s KimiK3Codec rejects it.
  • Muse Spark Open: The model’s Hugging Face repository is gated. To run a cookbook recipe on it, set HF_TOKEN to a token from an account that has access to the repository.
  • DeepSeek V4.1 Flash 1017: The cookbook has no codec for this model yet, so cookbook recipes can’t run on it. To train it, build samples by hand from the model’s chat template, and pass the effort as reasoning_effort. For data without reasoning, set enable_thinking=False. Otherwise, the template renders each reply after an empty reasoning block. When a conversation defines tools, the template keeps earlier reasoning.

Checkpoints and deployment

How you serve a trained model depends on its base model: How to read the columns:
  • Registers against: The serving base model that adapters from a LoRA session’s inference checkpoints register against. Not registered means the checkpoints aren’t uploaded to the model registry, so they can’t be deployed. Download the weights instead.
  • Checkpoint variants: The download variants of a LoRA session’s inference checkpoint. A model marked Adapter only has no merged variant, so download CHECKPOINT_VARIANT_ADAPTER.
  • Merged (--merge): Whether an adapter can deploy in merged mode, where each replica merges it into the base model when it starts. If a model marked No lists an FP4 preparation time, its merged models are served only in FP4, so adapters trained on it go through FP4 preparation instead.
  • Multi-LoRA: Whether adapters can attach to a multi-LoRA deployment of the serving base model.
  • FP4 preparation: The typical running time of an FP4 preparation job, excluding any wait for GPU capacity. Prepared models register against these serving base models:
    • GLM 5.3: zai-org/GLM-5.3-NVFP4.
    • GLM 5.3 Flash: zai-org/GLM-5.3-Flash-NVFP4.
    • Kimi K3: moonshotai/Kimi-K3, in the release’s own MXFP4 format.
    • MiniMax M3.1 Flash: MiniMaxAI/MiniMax-M3.1-Flash-NVFP4. The job only merges the adapter, without calibration or quantization, because the release already ships its routed experts in FP4.
Adapters trained on a single-session variant register against the same serving base model as the default model’s adapters, so they deploy the same way. GLM 5.3 adapters saved before October 9, 2026 record zai-org/GLM-5.3-NVFP4 as their base. They can’t attach to a multi-LoRA deployment, but FP4 preparation accepts them.

List supported models programmatically

To list the live B200 configurations:
The output has one block per model, like this:
To run a similar listing from a cookbook checkout, use uv run python scripts/supported_models.py.

Choose a GPU type

Set gpu_type="B200-SXM" inside ComputeConfig when you create a resource. The trainer and generator always run on the same GPU type. Check the model’s row in the catalog for the training modes available on B200.

Request a model

To request a model or config that isn’t in the catalog, contact your Together AI account team.