Frontis-MA1

Frontis-MA1 is a family of open-weight large language models post-trained to act as agents for machine learning engineering (MLE), released in late July 2026 by FrontisAI, a China-based research group formed by Horizon Research of Frontis.AI together with researchers at Tsinghua University.^[1][4] The flagship Frontis-MA1-35B is a 35-billion-parameter mixture-of-experts finetune of Alibaba's Qwen3.6-35B-A3B that activates roughly 3 billion parameters per token.^[2] Its technical report describes the model as an "AI4AI" system, an AI trained to improve the process of building AI, and labels it "Meta-evolution Agent, generation 1": the model is trained on the same four program-evolution operators, Draft, Improve, Debug, and Crossover, that its companion evolutionary search harness composes into a long-horizon loop at inference time, so that learning and search share a single action space grounded in real execution feedback.^[1] Frontis-MA1 was released together with OpenMLE, a full-stack open system for studying recursive self-improvement in machine learning engineering, under the group's broader OpenRSI initiative.^[3]

On the official 22-task Lite split of OpenAI's MLE-bench benchmark, with a fixed budget of 12 hours per task on a single NVIDIA RTX 4090 capped at 12 GB of VRAM, the paper reports that swapping the base Qwen3.6 model for Frontis-MA1-35B under an identical OpenMLE-Evo harness raises the Medal Average (any-medal rate) from 39.39% to 60.61%, and that the combined system reaches 71.21% with the enhanced OpenMLE-Evo-Max configuration, exceeding the paper's GPT-5.5 plus Codex baseline by 3.03 percentage points and approaching its GPT-5.6 Sol and Kimi K3 results.^[1] FrontisAI reports all headline numbers as properties of the combined model and harness system rather than of the model alone.^[2] The model weights and the OpenMLE code are licensed under CC BY-NC 4.0, which forbids commercial use.^[2][3]

Key facts

Scroll sideways for more →

Background: FrontisAI and OpenRSI

FrontisAI describes itself as working on "Scalable AI4AI & RSI"; its GitHub organization, created June 17, 2026, lists its location as China, and the paper's corresponding author, Kaiyan Zhang, uses a frontis.cn address.^[4][1] The paper credits two affiliations, Horizon Research at Frontis.AI and Tsinghua University, and names Junlin Yang and Che Jiang as project leaders among 24 authors.^[1]

OpenRSI is the group's open initiative for making "AI improving AI" an executable, measurable engineering problem. Its stated mechanism ladder runs from evolution through self-evolution and meta-evolution toward recursive self-improvement, and the group is explicit that it begins at the meta-evolution stage, "training the improver itself in bounded, executable domains," "without claiming that general recursive self-improvement has been solved."^[3] Before Frontis-MA1, the group released NatureBench (June 23, 2026), a benchmark testing whether coding agents can match the published state of the art of Nature-family papers, and a survey of self-improving agents (June 25, 2026).^[3][6]

The OpenMLE stack

Frontis-MA1 is both a product of the OpenMLE training stack and the model that drives its search loop.^[2] The stack, released in the FrontisAI/OpenRSI repository, has three main components:^[1][3]

• OpenMLE-Gym: verifiable machine learning engineering task environments with isolated execution, structured feedback, and task-specific evaluators; the paper says it unifies 5,758 tasks.^[1]
• OpenMLE-ERL (operator learning): execution-grounded supervised fine-tuning and online reinforcement learning that teach the model four reusable program-transformation operators: Draft (create an initial solution), Improve (refine a parent program using its score and execution evidence), Debug (repair failing code), and Crossover (recombine elements of two parent solutions).^[1][2]
• OpenMLE-Evo: a long-horizon evolutionary search harness that composes the four operators at inference time. OpenMLE-Evo-Max is an enhanced profile that adds cross-task experience priors distilled from public competition artifacts (with all MLE-bench-related sources excluded) and asynchronous multi-GPU search while keeping the same total sandbox-compute budget.^[1][3]

Because the same four operators define the action space during training and during search, the paper describes the arrangement as coupling learning and evolution in a single loop, and calls a model trained for this role a meta-evolution agent.^[1] FrontisAI also released two datasets on Hugging Face: OpenMLE-Tasks (task artifacts and construction pipelines) and OpenMLE-SFT-Traces (the supervised trajectories used to train the model).^[3][7]

Model and training

Frontis-MA1-35B inherits the Qwen3.6-35B-A3B architecture: a multimodal conditional-generation model with a hybrid-attention mixture-of-experts text backbone, 40 layers, 256 routed experts with 8 routed plus 1 shared expert active per token, and a native 262,144-token context window.^[2] The repository ships the base model's vision encoder and multi-token-prediction components unchanged; OpenMLE post-training and all reported evaluations are text and code only, and the model card states that they do not establish validated visual capability.^[2]

Post-training proceeded in two stages. Supervised fine-tuning used 26,259 released execution-grounded examples (17,245 full responses and 9,014 trajectory steps) constructed from programs executed in isolated MLE sandboxes, trained full-parameter in BF16 on 8 NVIDIA H200 GPUs for three epochs.^[2] Reinforcement learning then used online generation and evaluation in task sandboxes with the GSPO algorithm, an operator mixture of Draft 0.50, Improve 0.17, Debug 0.17, and Crossover 0.16, and a maximum response length of 24,576 tokens.^[2] The paper states that training data was deduplicated against all evaluation benchmarks.^[1]

A companion model, Frontis-MA1-30B, applies the same recipe to Qwen3-30B-A3B-Thinking-2507 to test whether the gain reproduces on a second backbone; it improves that base from 34.85% to 53.03% Medal Average under the same harness.^[1][2]

Benchmark results

MLE-Bench Lite

Evaluation uses the official 22-task MLE-Bench Lite split, the Low-complexity subset of MLE-bench's 75 Kaggle competitions, with three independent runs per configuration and a per-task budget of 12 hours on a single RTX 4090 capped at 12 GB VRAM, a smaller sandbox-compute budget than most reported MLE-bench evaluations.^[1] The controlled comparison holds the harness fixed and changes only the model:^[1][2]

Scroll sideways for more →

The first row-to-row gain (+21.22 percentage points) isolates the effect of post-training under a fixed harness; the second (+10.60 points) is a system-level search gain that the model card says should not be read as a pure model improvement.^[2] The paper states that the 71.21% system result exceeds its GPT-5.5 plus Codex baseline (68.18%) by 3.03 percentage points and is comparable to its GPT-5.6 Sol and Kimi K3 configurations, which both reach 72.73% on the final test with their respective harnesses.^[1]

Under the same fixed OpenMLE-Evo harness, the model card also reports results for other open-weight models: Kimi K2.6 (66.67%) and GLM-5.2 (62.12%) score above Frontis-MA1-35B's 60.61%, while MiniMax M3 (59.09%) and DeepSeek-V4-Flash (51.52%) score below it.^[2] The paper's claim is therefore not that Frontis-MA1-35B is the strongest model in the harness, but that execution-grounded post-training produces a large controlled gain over its own base at a much smaller parameter count.^[1][2]

Transfer to NatureBench Lite

To test transfer beyond Kaggle-style competition, the group evaluated a fixed 10-task subset of its own NatureBench, a 90-task benchmark of containerized problems distilled from peer-reviewed Nature-family papers, under a four-hour search budget per task with web search disabled.^[1][6] Holding the search adapter fixed, Frontis-MA1-35B raised Match-SOTA (matching the published result) from 50% to 70% of tasks and Surpass-SOTA from 20% to 30% over its base model; holding the model fixed, the OpenMLE-Evo adapter raised Match-SOTA from 20% to 50% compared with the original AIRA-Evo harness.^[1] The paper characterizes this ten-task study as evidence of focused transfer rather than a claim of general scientific autonomy.^[2]

Release and licensing

The Hugging Face repositories for Frontis-MA1-35B, Frontis-MA1-30B, and their GGUF quantizations were created on July 30, 2026, and the release was announced on July 31, 2026, alongside the OpenMLE code and datasets; the paper was posted to arXiv on July 30, 2026.^[5][3][1] The BF16 Transformers repository is the canonical release, with a Q4_K_M GGUF derivative for local deployment.^[2]

Original Frontis-MA1 and OpenRSI material, covering the model weights and the code repository, is licensed under CC BY-NC 4.0, an attribution-required, non-commercial license; the upstream Qwen Apache 2.0 notice is preserved separately.^[2][3] The release is therefore open-weight but not open source in the commonly used sense: commercial use is not granted.^[2] The model card also cautions that generated code may be incorrect, insecure, destructive, or expensive to execute, and should be run only inside isolated environments with explicit resource limits.^[2]

See also

• MLE-bench
• Harness (AI)
• Recursive self-improvement
• AutoML
• Qwen3.6
• Agentic AI

References

1. ^Yang, Junlin; Jiang, Che; Fu, Yu; et al. (Horizon Research, Frontis.AI and Tsinghua University). "Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering." arXiv:2607.28568, July 30, 2026. arxiv.org/...2607.28568
2. ^FrontisAI. "Frontis-MA1-35B" (model card). Hugging Face, July 2026. huggingface.co/...Frontis-MA1-35B
3. ^FrontisAI. "OpenRSI" (repository README). GitHub, July 31, 2026. github.com/...OpenRSI
4. ^FrontisAI organization profile. GitHub. github.com/FrontisAI
5. ^Hugging Face model API metadata, FrontisAI/Frontis-MA1-35B and related repositories (creation dates, base model, license tags). huggingface.co/...Frontis-MA1-35B
6. ^Wang, et al. "NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?" arXiv:2606.24530, June 23, 2026. arxiv.org/...2606.24530
7. ^FrontisAI. "OpenMLE-Tasks" and "OpenMLE-SFT-Traces" (datasets). Hugging Face, July 2026. huggingface.co/...OpenMLE-Tasks