Mantis

Mantis is an open-weights computer-use model checkpoint fine-tuned from Hcompany/Holo3-35B-A3B. It is trained on graded Mantis agent rollouts from Augur, using supervised fine-tuning over real browser/workflow traces rather than generic chat data.

This release includes both the PEFT LoRA adapter and a ready-to-serve merged.Q8_0.gguf artifact for llama.cpp-style serving.

Why This Is Better For Mantis

Mantis is not a general chat fine-tune. It is specialized for the slow improvement loop of a computer-use agent:

• Agent-native data: trained from Mantis rollouts with task context, model I/O, rewards, and action traces.
• Computer-use alignment: targets GUI/navigation behavior on realistic browser tasks instead of instruction-following only.
• Deployment-ready release: ships the small adapter plus a merged Q8_0 GGUF, so downstream serving stacks can either compose with the base model or run the merged artifact directly.
• Auditable provenance: checkpoint id sft-c3e0d799f432-f00fa0 ties this release to the training data/config hash used by the Mantis trainer registry.

The current internal frozen holdout gate did not establish a reliable promotion over the base model, so this page does not claim a benchmark win over Holo3. The value of this release is open access to the specialized Mantis adaptation, its reproducible training pipeline, and its serving artifact.

Files

• adapter_model.safetensors: PEFT LoRA adapter weights.
• adapter_config.json: PEFT adapter configuration.
• adapter.gguf: converted adapter artifact.
• merged.Q8_0.gguf: merged full-model Q8_0 GGUF for direct serving.
• tokenizer/processor files copied from the training artifact.
• training_args.bin: trainer metadata from the SFT run.

Intended Use

Use this checkpoint for research and development of GUI agents, browser automation agents, and Mantis-compatible computer-use systems.

This model may be useful when you need:

• a Holo3-derived checkpoint adapted to Mantis rollouts;
• an open adapter for further fine-tuning;
• a ready GGUF artifact for serving experiments;
• a transparent artifact from a champion/challenger training loop.

Limitations

• This model can make incorrect UI decisions and should not be allowed to take high-impact actions without supervision.
• The released checkpoint is specialized for Mantis-style workflows; behavior outside that domain may not improve over the base model.
• The internal gate found no reliable promotion over the base model on the frozen holdout available at release time.
• Computer-use agents can interact with external systems. Use sandboxing, allowlists, rate limits, and human approval for sensitive workflows.

Base Model And License

This model is fine-tuned from Hcompany/Holo3-35B-A3B, whose model card declares the Apache-2.0 license. This release is also published under Apache-2.0 and retains upstream attribution.

Training

• Base model: Hcompany/Holo3-35B-A3B
• Method: supervised fine-tuning with TRL
• Checkpoint id: sft-c3e0d799f432-f00fa0
• Data source: graded Mantis rollouts pulled from Augur
• Training stack: PEFT LoRA + TRL SFT

Citation

If you use the base model, cite Holo3:

@misc{hai2025holo3modelfamily,
  title={Holo3 - Open Foundation Models for Navigation and Computer Use Agents},
  author={H Company},
  year={2026},
  url={https://huggingface.co/Hcompany/Holo3-35B-A3B}
}

If you use the trainer stack, cite TRL:

@software{vonwerra2020trl,
  title={{TRL: Transformers Reinforcement Learning}},
  author={von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license={Apache-2.0},
  url={https://github.com/huggingface/trl},
  year={2020}
}