Skip to content

Model and algorithm support

Use the matrix to choose a model and training method, then follow a linked recipe or adapt the closest example. Support and recipe availability are separate: an existing model/algorithm core counts as supported even when configuration, launch scripts or data wiring still need adaptation.

Support matrix

  • Yes: the model adapter and algorithm support this combination.
  • Yes*: also supported through the existing core, with additional model/data wiring or runtime setup described below.
  • R: a reference recipe, configuration example or documented launch is available. The recipe sections distinguish these starting points and their scope.
  • —: the current adapters lack a required capability; configuration or data wiring alone is insufficient. Unlisted combinations may still be adaptable.

Support applies to compatible checkpoints on the selected backend; it does not imply that every checkpoint in a family has been tested. R applies to the linked model/task, rather than the whole family.

Model family / backend SFT Flow-DPO GRPO / Dr-GRPO / GSPO Token PPO Joint PPO OPD DiffusionNFT
Cosmos3 generator / OmniMoT Yes R Yes R — — — — Yes R: image
HF Qwen text Yes — Yes R: GRPO Yes R — Yes R —
Other compatible HF causal decoders Yes* — Yes* Yes* — Yes* —
HF dense Qwen3.5/3.8 (qwen3_5) Yes R — Yes* Yes* — Yes* —
HF dense Qwen3-VL Yes — Yes Yes — Yes —
Other compatible HF dense VLMs Yes* — Yes* Yes* — Yes* —
Native Qwen3-VL / Qwen3-VL-MoE Yes R — Yes Yes R — Yes R —
GR00T-N1.7 Yes* — — — Yes R — —
π0.5 — — — — Yes R — —
Alpamayo / AlpaGym — — Yes* — — — —

Scope and adaptation notes

  • HF SFT: the default conversation-SFT launcher covers Qwen text (qwen2, qwen2_moe, qwen3, qwen3_moe), dense Qwen3.5/3.8 and Qwen3-VL. Other compatible causal decoders and dense VLMs can reuse the CE objective and trainer with a model-specific tokenizer/template, processor, packer and launcher binding. Qwen3-VL-MoE uses the native backend.
  • Online HF methods: the checkpoint must work with the training backend, rollout/teacher runtime and live-weight loading; PPO also needs a compatible critic. Dense Qwen3.5/3.8 requires a runtime with model support. Generic HF VLM packing covers text and images; native Qwen3-VL has a separate video packer. HF VLM OPD uses the same-token (shared-tokenizer) path; cross-token multimodal OPD uses native Qwen3-VL and currently accepts images.
  • GR00T SFT: supervised action-head flow matching is implemented, with the backbone frozen. An offline data source and training loop must be assembled around that core; a public end-to-end customer SFT recipe is not provided.
  • Algorithm variants: native Qwen3-VL GRPO uses the clipped surrogate. Alpamayo's AlpaGym integration provides closed-loop GRPO-like training; its cell does not include Dr-GRPO or GSPO.
  • Diffusion methods: Flow-DPO is preference optimization for the Cosmos3 flow adapter. DiffusionNFT implements image and video paths; the linked public recipe covers images. GR00T and π0.5 use different action-policy interfaces.
  • Cosmos3 Reasoner: export as Qwen3-VL and use that row for token-based methods; see checkpoint conversion.

Supervised training

Use the native Cosmos3 trainer for Cosmos3 recipes, or the Posttrain conversation-SFT runtime for Qwen and compatible HF models. GR00T uses a separate flow-matching objective.

Model / workflow Starting point Scope
Cosmos3-Nano / Edge / Super Training guide, recipe catalogue Generator and Reasoner recipes; full training and LoRA availability depend on the recipe.
Cosmos3-Nano robot policies DROID policy, forward dynamics, LIBERO policy Cosmos3 observation-to-action and action-conditioned generation recipes.
HF Qwen conversation SFT SFT guide, Qwen3.8-27B multimodal recipe Full-parameter SFT with local/torchrun FSDP; this path does not provide LoRA or sequence packing.
Native Qwen3-VL RoboVQA Qwen3-VL-8B recipe Recipe includes held-out task measurements.
Native Qwen3-VL-MoE SFT configuration example Read its precision, optimizer and parameter-update constraints before adapting it.
GR00T-N1.7 Supervised model/objective factories, training parity harness Action-head SFT building blocks; for a customer workflow, start with Isaac-GR00T.

For other compatible HF families, follow the adaptation guides. The generic SFT core is exercised by GPT-2 loss and optimizer tests, even though the default launcher selects Qwen families.

Preference optimization, RL and on-policy distillation

The public run guides provide the launch environment and recorded validation results. Configuration examples are additional starting points. Short optimizer runs establish that a path executes; assess quality on your task separately.

Method / model Starting point Scope
Flow-DPO / Cosmos3-Nano T2I recipe, DPO guide Public optimizer smoke uses synthetic preference pairs.
GRPO / Qwen2.5-3B GSM8K recipe, GRPO guide Public launch recipe; compatible models can reuse the token-policy path.
Dr-GRPO / GSPO Dr-GRPO configuration, GSPO configuration Configuration examples; adapt the public GRPO launch as described in its guide.
Token PPO / Qwen3-0.6B Four-GPU recipe, PPO guide Public policy/critic update recipe.
Token PPO / native Qwen3-VL-MoE Qwen3-VL-30B-A3B configuration Configuration example using native policy and value builders.
Joint PPO / GR00T-N1.7 Public guide, four-GPU recipe LIBERO-Spatial checkpoint, embodiment and simulator contract.
Joint PPO / π0.5 Public guide, four-GPU recipe LIBERO-Spatial; model weights and observation processing differ from GR00T.
OPD / text and native Qwen3-VL Countdown text recipe, OPD guide Guide also includes a native multimodal launch and explains teacher modes.
DiffusionNFT / Cosmos3-Nano T2I recipe, NFT guide Public image recipe with a toy reward; video requires its own configuration and reward service.
Closed-loop RL / Alpamayo AlpaGym integration entry point Requires the external AlpaGym runtime, policy bundle and simulator; no standalone public recipe listed here.

For Cosmos3 teacher/student diffusion-distillation curricula, see Choose a path.

GR00T: choosing a repository and handing off a checkpoint

The repositories currently offer complementary workflows:

  1. Customer demonstrations → SFT: start with Isaac-GR00T fine-tuning for dataset preparation, training, evaluation and deployment. Cosmos Framework also supports supervised action-head training, but does not provide the equivalent end-to-end SFT recipe.
  2. Reward-based improvement: start with the Cosmos Framework GR00T Joint PPO guide. To use a customer SFT checkpoint, match its embodiment, cameras/state, action layout, horizon, processor and normalization statistics to the training and simulator configuration.
  3. Evaluation/deployment after RL: the GR00T exporter replaces a matching base checkpoint's action head and requires matching keys and shapes. Check compatibility with PPO's added value head before exporting. Preserve processor, statistics and configuration artifacts, then test the policy in the intended Isaac-GR00T runtime.

The public LIBERO recipe demonstrates its named checkpoint/environment combination. A customer SFT → CF RL → deployment handoff needs the compatibility checks above; the recipe does not establish that full round trip for arbitrary checkpoints.

Adding or qualifying another combination

Unlisted combinations can often be adapted from existing examples. Start from the closest R example with the same model class and training signal. Compatible models can reuse the objective and trainer while you connect the checkpoint, tokenizer/template, processor and dataset.

What changes Guide / reference
Tokenizer/chat template, media inputs or architecture-specific packing Custom DataPacker
Dataset format or a run-local data factory Custom datasets and DataPackers
Loss, reduction or algorithm variant Objective customization
Task scorer or reward service Reward providers
Robot embodiment, observation/action mapping or simulator Simulator integration
Model forwards, value heads or weight layout Model capability protocols, model build specification, extending beyond an objective

Bind the model and data first. For online methods, also connect rollout/scoring, live-weight transfer and any teacher, reference or value model. Run preflight checks, then a short training/resume/export smoke before evaluating task quality. If a required model-loading or loss capability is missing, add the corresponding adapter alongside the configuration. A working adaptation can become a new reference recipe.

Implementation references behind the matrix