kehong/TinyAstral-weights

TinyAstral weights

Paper: arXiv:2610.41216

_{85-second overview (with sound). Also on the project homepage.}

TinyAstral students are small vision-language models that act as the decision model inside the TinyAstral harness $H^*$. They were distilled from a frozen ~180B open-weight teacher (Qwen3.8-Flash-Next) that ran in the same harness. The harness supplies perception, geometry and motion through tools. The model chooses the next tool call from camera images, tool results and its own short reasoning.

Folder	Base model	Parameters	Files
4b/	Qwen3.5-4B	4B	3 safetensors shards (bf16), 10.3 GB
2b/	Qwen3.5-2B	2B	2 safetensors shards (bf16), 5.4 GB
0.8b/	Qwen3.5-0.8B	0.8B	1 safetensors file (bf16), 2.2 GB

Each folder is a complete Hugging Face checkpoint (config, tokenizer, chat template, processor config).

These models are not standalone vision-language-action policies. They emit harness tool calls, and they need the TinyAstral harness and its perception service to act. Use them with the TinyAstral code release.

Training

• Stage 1. The teacher's harness was improved by recursive self-improvement on development and selection seeds, then frozen as $H^*$.
• Stage 2. The frozen teacher with $H^*$ collected trajectories on training seeds of all 50 RoboTwin 2.0 tasks, the 7 seen Guava tasks and LIBERO-90. This gave main-episode decisions plus recovery branches started from perturbed replays. The data also includes supervision built from real-robot datasets: verifier, planner and next-call samples. In total there are 83,753 decision samples, about half of them simulated.
• SFT. Full fine-tuning of the language model with the vision encoder and projector frozen. Learning rate 1e-5, cosine decay, 5% warm-up, global batch 32, maximum length 12,288 tokens, 2 epochs. Loss is applied to the final assistant turn only. All three sizes use the same data and recipe.

Results (success %, same harness $H^*$, strict protocol)

Training benchmarks use strict success: an autonomous done and terminal success. Zero-shot benchmarks use their official measures.

Model	RoboTwin-50 (500 ep.)	Guava (450)	LIBERO-PRO (300)	SimplerEnv Google VM	SimplerEnv WidowX	VLABench (500)	EB-Manipulation (228)
Teacher (Qwen3.8-Flash-Next, ~180B)	63.6	94.7	80.0	69.2	92.7	46.2	32.0
Qwen3.5-4B, untrained	41.2	75.1	36.3	56.6	86.5	27.6	21.9
TinyAstral-4B	62.8	93.6	78.3	66.6	88.5	41.2	27.6
TinyAstral-2B	59.8	88.0	69.0	65.4	88.5	42.0	25.0
TinyAstral-0.8B	61.8	93.8	54.7	64.0	87.5	40.4	24.6

SimplerEnv, VLABench and EB-Manipulation are zero-shot: no training data or harness tuning came from them.

Serving

GPU (vLLM). One GPU, bf16:

vllm serve <local path>/4b --served-model-name tinyastral-4b --dtype bfloat16 \
  --max-model-len 32768 --limit-mm-per-prompt '{"image": 8}'

Point the harness at the OpenAI-compatible endpoint. Thinking should stay enabled, as in training. The TinyAstral harness sends chat_template_kwargs: {"enable_thinking": true} with every request. If you call the models directly, pass it yourself: the 2B and 0.8B chat templates (inherited from the Qwen3.5 small models) default to thinking off, while 4B defaults to on. The checkpoints contain no multi-token-prediction weights, so do not enable MTP speculative decoding.

Huawei Ascend Atlas A2 (SGLang). The code release includes scripts for serving on Huawei Ascend Atlas A2 with SGLang.

Intended use and limitations

• Research on embodied agents with tool-based harnesses, in simulation and on real robots through the TinyAstral harness.
• The models assume the TinyAstral tool interface and prompt format. Other tool layers will need adaptation.
• Not evaluated for safety-critical deployment.

License

The TinyAstral weights are released under CC BY-NC-SA 4.0 (non-commercial, share-alike), as a whole. They were trained on the TinyAstral-SFT dataset, which is CC BY-NC-SA 4.0 because it includes samples derived from AgiBot World (CC BY-NC-SA 4.0) and RH20T (partly CC BY-NC 4.0); the weights carry the same terms.

• Base models: Qwen3.5-0.8B, Qwen3.5-2B and Qwen3.5-4B, Apache-2.0.
• Teacher outputs in the training data were generated with Qwen3.8-Flash-Next and are subject to its license.
• Real-robot data sources and their licenses are listed in the TinyAstral-SFT dataset card: DROID, RT-1 (Fractal), Bridge V2 and Berkeley UR5 (CC BY 4.0); AgiBot World (CC BY-NC-SA 4.0); RH20T (CC BY-SA 4.0 / CC BY-NC 4.0 subsets).
• The TinyAstral code (harness, adapters, serving, training recipe) is Apache-2.0.

Citation

@article{gong2026tinyastral,
  title         = {TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents},
  author        = {Gong, Kehong and Zheng, Xiaoxu and Hou, Guanli and Lee, Seunghun and Guo, Mengqi and Tan, Runjia and Wen, Zhengyu and Wang, Qi and He, Xiaoyu and Zhang, Mingyuan and Mi, Michael Bi},
  journal       = {arXiv preprint arXiv:2610.41216},
  year          = {2026},
  eprint        = {2610.41216},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.41216}
}