Frontis-MA1: Open 35B model toward recursive self-improvement

14 hours ago

OpenMLE is an open full-stack system for recursive self-improvement (RSI) in machine learning engineering (MLE), including environments (Gym), operator learning (RL), and long-horizon search (Evo).
Frontis-MA1 (35B) is post-trained as a meta-evolution agent, using four atomic operators (Draft, Improve, Debug, Crossover) trained via execution-grounded SFT and RL.
On MLE-Bench Lite, Frontis-MA1 improves Medal Average from 39.39% to 71.21% with OpenMLE-Evo-Max, surpassing GPT-5.5 + Codex and approaching GPT-5.6 Sol.
Both the trained model and the search framework transfer to held-out NatureBench Lite, raising Match-SOTA from 50% to 70% and 20% to 50% respectively.
The full stack—weights, gym, sandbox, training, search, and evaluation—is open-sourced for reproducible AI4AI research towards RSI.

Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with 8B active parameters per token and a 1M-token context window, developed in under nine weeks.
It achieves competitive scores against much larger models on long-horizon coding benchmarks, scoring 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1.
All evaluation trajectories are publicly released at trajectories.poolside.ai for transparency and reproducibility.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

AMD introduces Instella-MoE, a fully open Mixture-of-Experts language model with 16 billion total parameters and 2.8 billion active parameters per token.
The model is trained from scratch on AMD Instinct MI300X and MI325X GPUs using the AMD ROCm software stack, incorporating architectural innovations like Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective connectivity.
Instella-MoE undergoes a multi-stage training pipeline including pre-training, mid-training, long-context extension, supervised fine-tuning, direct preference optimization, and reinforcement learning, with all artifacts openly released.

Introduction of Ornith-1.0, a self-improving open-source model family for agentic coding tasks.
Model variants include 9B Dense, 31B Dense, 35B MoE, and 397B MoE, built on pretrained Gemma 4 and Qwen 3.5.
Key innovation: self-improving training framework where the model learns to generate both solution rollouts and task-specific harnesses.

Solar Open 2 is an open-weight foundation model optimized for agentic use, designed to serve as a base model for agents in real work environments.
It uses a Mixture-of-Experts (MoE) architecture with 250B total parameters but activates only 15B per token, improving both capability and inference efficiency.
The model supports a 1M-token context window and runs on two NVIDIA H200 GPUs with quantization, making it practical for enterprise deployment.

Laguna S 2.1 is a 118B-parameter Mixture-of-Experts model with 8B active parameters per token and a 1M-token context window, developed in under nine weeks.
It achieves competitive scores against much larger models on long-horizon coding benchmarks, scoring 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE v1.1.
All evaluation trajectories are publicly released at trajectories.poolside.ai for transparency and reproducibility.

Inkling-Small is a general-purpose multimodal model that accepts text, image, and audio inputs and generates text outputs.
It can be used with libraries like Transformers, vLLM, SGLang, and Docker, as well as through APIs and third-party providers.
The model uses a 42-layer decoder-only transformer with a sparse Mixture-of-Experts (MoE) architecture, totaling 276B parameters but activating only 12B per token.

AMD introduces Instella-MoE, a fully open Mixture-of-Experts language model with 16 billion total parameters and 2.8 billion active parameters per token.
The model is trained from scratch on AMD Instinct MI300X and MI325X GPUs using the AMD ROCm software stack, incorporating architectural innovations like Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective connectivity.
Instella-MoE undergoes a multi-stage training pipeline including pre-training, mid-training, long-context extension, supervised fine-tuning, direct preference optimization, and reinforcement learning, with all artifacts openly released.

Introduction of Ornith-1.0, a self-improving open-source model family for agentic coding tasks.
Model variants include 9B Dense, 31B Dense, 35B MoE, and 397B MoE, built on pretrained Gemma 4 and Qwen 3.5.
Key innovation: self-improving training framework where the model learns to generate both solution rollouts and task-specific harnesses.

Solar Open 2 is an open-weight foundation model optimized for agentic use, designed to serve as a base model for agents in real work environments.
It uses a Mixture-of-Experts (MoE) architecture with 250B total parameters but activates only 15B per token, improving both capability and inference efficiency.
The model supports a 1M-token context window and runs on two NVIDIA H200 GPUs with quantization, making it practical for enterprise deployment.

Hasty Briefsbeta