README.md

---
license: cc-by-nc-4.0
base_model: FrontisAI/Frontis-MA1-35B
base_model_relation: quantized
library_name: llama.cpp
pipeline_tag: image-text-to-text
tags:
  - openmle
  - frontis-ma1
  - gguf
  - q4-k-m
  - local-inference
  - multimodal
  - moe
  - coding
---

# Frontis-MA1-35B-GGUF

<p align="center">
  <a href="https://arxiv.org/abs/2607.28568">📄 Paper</a>
  &nbsp;•&nbsp;
  <a href="https://frontisai.github.io/OpenRSI/">🌐 Project</a>
  &nbsp;•&nbsp;
  <a href="https://github.com/FrontisAI/OpenRSI">💻 Code</a>
  &nbsp;•&nbsp;
  <a href="https://huggingface.co/collections/FrontisAI/frontis-ma1">🤗 Models</a>
  &nbsp;•&nbsp;
  <a href="https://huggingface.co/datasets/FrontisAI/OpenMLE-Tasks">🧩 Tasks</a>
  &nbsp;•&nbsp;
  <a href="https://huggingface.co/datasets/FrontisAI/OpenMLE-SFT-Traces">📚 SFT Traces</a>
</p>

This repository is the official local-deployment derivative of [Frontis-MA1-35B](https://huggingface.co/FrontisAI/Frontis-MA1-35B). It contains one `Q4_K_M` language-model file and the F16 multimodal projector required for image input with `llama.cpp`.

It accompanies the paper [*Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering*](https://arxiv.org/abs/2607.28568) and the [OpenRSI code release](https://github.com/FrontisAI/OpenRSI).

## Files

| File | Size | Purpose |
| --- | ---: | --- |
| `Frontis-MA1-35B-Q4_K_M.gguf` | 19.71 GiB | Q4_K_M language model |
| `mmproj-Frontis-MA1-35B-F16.gguf` | 857.62 MiB | F16 vision encoder/projector |
| `checksums.txt` | — | SHA-256 integrity manifest |

Only this deployment combination is published intentionally. The canonical BF16 Transformers weights remain in the base repository. This GGUF derivative does not publish a separate MTP draft-model variant.

## Text and code quickstart

Tested conversion and inference tool: [`llama.cpp` b9637](https://github.com/ggml-org/llama.cpp/releases/tag/b9637), commit `aedb2a5e9ca3d4064148bbb919e0ddc0c1b70ab3`.

```bash
llama-cli \
  -m ./Frontis-MA1-35B-Q4_K_M.gguf \
  -ngl all \
  -c 32768 \
  -n 1024 \
  -cnv -st --simple-io \
  -p "Build a strong tabular classification baseline and explain the validation design."
```

## Image quickstart

```bash
llama-cli \
  -m ./Frontis-MA1-35B-Q4_K_M.gguf \
  -mm ./mmproj-Frontis-MA1-35B-F16.gguf \
  --image ./example.jpg \
  -ngl all \
  -c 32768 \
  -n 512 \
  -cnv -st --simple-io \
  -p "Describe the image and identify information relevant to an ML workflow."
```

Reduce `-c` when memory is limited. On systems that cannot offload all layers, set `-ngl` to a smaller value or let `llama.cpp` choose automatically.

## Release validation

Both final files passed SHA-256 verification and complete GGUF structure reads (733 language-model tensors and 334 projector tensors). The release also passed two real `llama-cli` smokes with full GPU offload on one NVIDIA H200: text generation from the Q4 file, and image-conditioned generation using the Q4 file with the F16 projector. These checks validate the release artifacts and command paths; they are not consumer-hardware speed benchmarks.

## Component and evaluation scope

- The language-model weights are the OpenMLE post-trained Frontis-MA1-35B weights.
- The vision encoder/projector is inherited unchanged from Qwen3.6-35B-A3B and converted to F16 GGUF.
- OpenMLE post-training and the reported evaluations are text/code-only; they do not establish improved or fully validated visual capability.
- `Q4_K_M` is lossy. Use the BF16 repository when maximum fidelity or paper-result reproduction is required.
- The paper's reported scores measure the canonical model with the OpenMLE-Evo harness, not GGUF one-shot generation.

Generated code may be incorrect or unsafe. Execute it only in an isolated environment with explicit resource limits.

## Paper result

The canonical BF16 model reaches **60.61% Medal Average** and **0.7647 Human Rank** with OpenMLE-Evo on the official 22-task MLE-Bench Lite split, compared with 39.39% and 0.5828 for its base model under the same harness. With OpenMLE-Evo-Max, the complete BF16 model–harness system reaches 71.21% and 0.8126. These are BF16 system results, not GGUF one-shot scores.

## License

Original Frontis-MA1 material is released under [CC BY-NC 4.0](LICENSE) for attribution-required, non-commercial use. Commercial use is not granted. The upstream Qwen Apache License 2.0 notice is preserved in [LICENSE-UPSTREAM-APACHE-2.0](LICENSE-UPSTREAM-APACHE-2.0) and [NOTICE](NOTICE).