# 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](https://github.com/NVIDIA/cosmos-framework/blob/main/examples/README.md) | Yes [R](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/dpo_t2i_public.py) | — | — | — | — | Yes [R: image](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/config/experiment/nft/cosmos3_nft_t2i_public_4gpu.py) | | HF Qwen text | Yes | — | Yes [R: GRPO](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/grpo_qwen3b_gsm8k_public_4gpu.py) | Yes [R](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/ppo_qwen06b_public_4gpu.py) | — | Yes [R](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/opd_qwen1p5b_countdown_public_4gpu.py) | — | | Other compatible HF causal decoders | Yes* | — | Yes* | Yes* | — | Yes* | — | | HF dense Qwen3.5/3.8 (`qwen3_5`) | Yes [R](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/sft_qwen3p8_27b_multimodal.py) | — | 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](#supervised-training) | — | Yes | Yes [R](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/ppo_qwen3vl30b_a3b.py) | — | Yes [R](../posttrain/guides/opd.md) | — | | GR00T-N1.7 | Yes* | — | — | — | Yes [R](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/gr00t_n1d7_libero_public_4gpu.py) | — | — | | π0.5 | — | — | — | — | Yes [R](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/pi05_libero_spatial_public_4gpu.py) | — | — | | 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](choose_path.md#rl). ## 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](../training.md), [recipe catalogue](https://github.com/NVIDIA/cosmos-framework/blob/main/examples/README.md) | Generator and Reasoner recipes; full training and LoRA availability depend on the recipe. | | Cosmos3-Nano robot policies | [DROID policy](../action_policy_droid_posttrain.md), [forward dynamics](../action_fd_droid_posttrain.md), [LIBERO policy](../action_policy_libero_posttrain.md) | Cosmos3 observation-to-action and action-conditioned generation recipes. | | HF Qwen conversation SFT | [SFT guide](https://github.com/NVIDIA/cosmos-framework/blob/main/docs/posttrain/guides/sft.md), [Qwen3.8-27B multimodal recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/sft_qwen3p8_27b_multimodal.py) | Full-parameter SFT with local/torchrun FSDP; this path does not provide LoRA or sequence packing. | | Native Qwen3-VL | [RoboVQA Qwen3-VL-8B recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/sft_qwen3_vl_8b_robovqa.py) | Recipe includes held-out task measurements. | | Native Qwen3-VL-MoE | [SFT configuration example](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/sft_qwen3_vl_moe_native.py) | Read its precision, optimizer and parameter-update constraints before adapting it. | | GR00T-N1.7 | [Supervised model/objective factories](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/gr00t_n1d7_droid.py), [training parity harness](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/tests/gr00t_train_parity_cli.py) | Action-head SFT building blocks; for a customer workflow, start with [Isaac-GR00T](https://github.com/NVIDIA/Isaac-GR00T#fine-tuning). | For other compatible HF families, follow the [adaptation guides](#adding-or-qualifying-another-combination). The generic SFT core is exercised by [GPT-2 loss and optimizer tests](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/tests/test_algorithm_sft.py), 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](../posttrain/index.md#qualified-commands). 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](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/dpo_t2i_public.py), [DPO guide](../posttrain/guides/dpo.md) | Public optimizer smoke uses synthetic preference pairs. | | GRPO / Qwen2.5-3B | [GSM8K recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/grpo_qwen3b_gsm8k_public_4gpu.py), [GRPO guide](../posttrain/guides/grpo.md) | Public launch recipe; compatible models can reuse the token-policy path. | | Dr-GRPO / GSPO | [Dr-GRPO configuration](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/dr_grpo_qwen3b_gsm8k.py), [GSPO configuration](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/gspo_qwen3b_gsm8k.py) | Configuration examples; adapt the public GRPO launch as described in its guide. | | Token PPO / Qwen3-0.6B | [Four-GPU recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/ppo_qwen06b_public_4gpu.py), [PPO guide](../posttrain/guides/ppo.md) | Public policy/critic update recipe. | | Token PPO / native Qwen3-VL-MoE | [Qwen3-VL-30B-A3B configuration](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/ppo_qwen3vl30b_a3b.py) | Configuration example using native policy and value builders. | | Joint PPO / GR00T-N1.7 | [Public guide](https://github.com/NVIDIA/cosmos-framework/blob/main/docs/posttrain/guides/joint-ppo-gr00t-public.md), [four-GPU recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/gr00t_n1d7_libero_public_4gpu.py) | LIBERO-Spatial checkpoint, embodiment and simulator contract. | | Joint PPO / π0.5 | [Public guide](https://github.com/NVIDIA/cosmos-framework/blob/main/docs/posttrain/guides/joint-ppo-pi05-public.md), [four-GPU recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/pi05_libero_spatial_public_4gpu.py) | LIBERO-Spatial; model weights and observation processing differ from GR00T. | | OPD / text and native Qwen3-VL | [Countdown text recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/configs/opd_qwen1p5b_countdown_public_4gpu.py), [OPD guide](../posttrain/guides/opd.md) | Guide also includes a native multimodal launch and explains teacher modes. | | DiffusionNFT / Cosmos3-Nano | [T2I recipe](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/config/experiment/nft/cosmos3_nft_t2i_public_4gpu.py), [NFT guide](../posttrain/guides/nft.md) | Public image recipe with a toy reward; video requires its own configuration and reward service. | | Closed-loop RL / Alpamayo | [AlpaGym integration entry point](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/entrypoints/alpagym_clrl.py) | 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](choose_path.md#distillation). ## 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](https://github.com/NVIDIA/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](https://github.com/NVIDIA/cosmos-framework/blob/main/docs/posttrain/guides/joint-ppo-gr00t-public.md). 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](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/scripts/export_gr00t_checkpoint.py) 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](../posttrain/howto/custom-datapacker.md#adding-a-packer-for-a-new-architecture) | | Dataset format or a run-local data factory | [Custom datasets and DataPackers](../posttrain/howto/custom-data.md) | | Loss, reduction or algorithm variant | [Objective customization](../posttrain/howto/custom-objective.md) | | Task scorer or reward service | [Reward providers](../posttrain/howto/custom-reward.md) | | Robot embodiment, observation/action mapping or simulator | [Simulator integration](https://github.com/NVIDIA/cosmos-framework/blob/main/docs/posttrain/howto/sim-integration.md) | | Model forwards, value heads or weight layout | [Model capability protocols](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/protocol.py), [model build specification](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/config/model.py), [extending beyond an objective](../posttrain/howto/custom-objective.md#when-objective-customization-is-not-enough) | 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](../posttrain/howto/preflight.md), 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 - **Generic SFT:** the [CE objective](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/algorithm/sft.py) works with compatible logits and SFT packing, as exercised by the [GPT-2 tests](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/tests/test_algorithm_sft.py). The default [SFT entry point](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/entrypoints/sft.py) and [packer builder](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/data/packer/sft.py) select a narrower set of Qwen families. - **HF token methods:** [HFModel](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/hf.py) provides causal/image-text forwards and loading. [Packer selection](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/data/packer/__init__.py), the [HF VLM packer](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/data/packer/hf_vlm.py) and [GRPO assembly](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/entrypoints/grpo.py) connect these to common objectives. - **Native Qwen3-VL:** [dense policy/value builders](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/qwen3_vl.py), [MoE policy](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/qwen3_vl_moe.py) and [media packing](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/data/packer/qwen3_vl.py) provide the separate native integration for `qwen3_vl` and `qwen3_vl_moe`. - **PPO and OPD roles:** [PPO assembly](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/entrypoints/ppo.py) binds policy, reference and [critic](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/value.py). [OPD assembly](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/entrypoints/opd.py) and [teacher/objective builders](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/config/build.py) select teacher modes; DPCA uses the native backend. - **Diffusion methods:** [DPO](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/entrypoints/dpo.py) and [NFT](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/entrypoints/nft.py) assemble the [Cosmos3 flow adapter](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/cosmos3.py). NFT includes video-specific preflight, sampling, decoding and reward wiring. - **Robot policies:** [GR00T](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/gr00t/policy.py) implements supervised flow matching for [flow-BC](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/algorithm/flow_bc.py) as well as decision-PPO methods. The current [π0.5 adapter](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/pi05/policy.py) provides decision PPO and lacks the supervised loss/input path; it needs additional implementation for SFT in CF. - **Online weights and checkpoints:** the [weight-transfer plane](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/comm/weight_transfer.py) and [vLLM receiver](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/roles/rollout/vllm_weights.py) must load the selected checkpoint's live tensors. [Checkpoint lifecycle](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/control/checkpointing.py), [HF export](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/model/hf.py) and [native SFT materialization](https://github.com/NVIDIA/cosmos-framework/blob/main/cosmos_framework/posttrain/checkpoint/tools/materialize_hf.py) cover different save/export paths; training resume and deployment export have distinct formats.