IQuest-Q1, a sparse Mixture-of-Experts agentic foundation model with 320B total parameters (15B activated), was released as an open-source model for agentic CLI and coding systems. Training combined Stable SplitNorm, Muon, and muP pre-training with multi-scaffold post-training across SFT, domain-specialized RL, and MOPD. The release targets community research and development.
IQuest released IQuest-Q1 on Hugging Face, a Mixture-of-Experts model targeting agentic coding, reasoning, and multi-step tool use. The model has roughly 320B total parameters with an estimated 15B activated per token, making it a sparse MoE design aimed at efficient inference for agent workloads.
IQuest released IQuest-Q1 on Hugging Face, a Mixture-of-Experts model targeting agentic coding, reasoning, and multi-step tool use. The model has roughly 320B total parameters with an estimated 15B activated per token, making it a sparse MoE design aimed at efficient inference for agent workloads.
A new arXiv paper introduces T1, a 122B-parameter Mixture-of-Experts model trained with reinforcement learning to operate a real shell in a cloud sandbox on long-horizon terminal tasks, with tasks lasting up to 300+ tool-call turns. The training recipe combines warm-started actor-critic training, dense process rewards from task verifiers, TITO token-level training with drift repair, and rollout routing replay for MoE stability. T1 reaches 64.0% on Terminal-Bench 2.1 and 27.9% on Long-Horizon Terminal Bench, surpassing GPT-5.4 and GLM-5.1 on the latter.
A new arXiv paper introduces T1, a 122B-parameter Mixture-of-Experts model trained with reinforcement learning to operate a real shell in a cloud sandbox on long-horizon terminal tasks, with tasks lasting up to 300+ tool-call turns. The training recipe combines warm-started actor-critic training, dense process rewards from task verifiers, TITO token-level training with drift repair, and rollout routing replay for MoE stability. T1 reaches 64.0% on Terminal-Bench 2.1 and 27.9% on Long-Horizon Terminal Bench, surpassing GPT-5.4 and GLM-5.1 on the latter.
Axion 1.0, a fully open agent foundation model, launched with weights, training data, code, and an arXiv paper. It uses a three-stage recipe — two SFT stages then reinforcement learning — without mid-training, DPO, or on-policy distillation, and reportedly matches leading commercial models on several agentic benchmarks. The complete data construction, filtering, verification, and training pipeline is released for reproduction.
Axion 1.0, a fully open agent foundation model, launched with weights, training data, code, and an arXiv paper. It uses a three-stage recipe — two SFT stages then reinforcement learning — without mid-training, DPO, or on-policy distillation, and reportedly matches leading commercial models on several agentic benchmarks. The complete data construction, filtering, verification, and training pipeline is released for reproduction.
A new arXiv technical report introduces Instella-MoE, a fully open Mixture-of-Experts language model with 16 billion total and 2.8 billion active parameters per token, trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. The work pairs sparse MoE design with architectural and system-level innovations aimed at efficient large-scale training and inference. It matters to practitioners as a fully open reference for MoE training outside the NVIDIA/CUDA ecosystem.
A new arXiv technical report introduces Instella-MoE, a fully open Mixture-of-Experts language model with 16 billion total and 2.8 billion active parameters per token, trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. The work pairs sparse MoE design with architectural and system-level innovations aimed at efficient large-scale training and inference. It matters to practitioners as a fully open reference for MoE training outside the NVIDIA/CUDA ecosystem.
The Solar Open 2 technical report describes a 250B-A15B Mixture-of-Experts language model designed for long-horizon agentic tasks. It achieves a 1M-token context window using a hybrid attention stack and efficient training methods.
The Solar Open 2 technical report describes a 250B-A15B Mixture-of-Experts language model designed for long-horizon agentic tasks. It achieves a 1M-token context window using a hybrid attention stack and efficient training methods.