Tracked topic
… Our inference kernel is available at: https://github.com/MiniMax-AI/MSA. A production-grade natively multimodal model powered by MSA has been publicly released at: https://huggingface.co/MiniMaxAI/MiniMax-M3. …
… Generated by Qwen/Qwen2.5-Coder-32B-Instruct We present MaxProof, a population-level test-time scaling framework for competition-level mathematical proof in the MiniMax-M3 series . …
… While we have only evaluated SparDA on 8B models, we hope it can shed light on inference optimization for larger frontier models with sparse attention capability, including DeepSeek-V4, GLM-5.1, and MiniMax M3. …
… The following papers were recommended by the Semantic Scholar API Terminal-World: Scaling Terminal-Agent Environments via Agent Skills 2026 TerminalWorld: Benchmarking Agents on Real-World Terminal Tasks 2026 Toward Scalable Terminal Task Synthesis via Skill Graphs 2026 The MiniMax-M2 Series: Mini … …
… View arXiv page View PDF GitHub 4 Add to collection Community We also evaluate the latest models Gemini-Omni-Flash, HappyHorse-1.1, and MiniMax-H3. …
… Quantitative consequences include: the minimax optimal generalization rate O n^{-1/ d+2 } for fixed t; adversarial robustness controlled by varepsilon; backpropagation as the co-state equation of the Hamiltonian system for residual networks Pontryagin Maximum Principle ; scaling exponents consisten… …
… The following papers were recommended by the Semantic Scholar API Laguna M.1/XS.2 Technical Report 2026 ZAYA1-8B Technical Report 2026 HEBATRON: A Hebrew-Specialized Open-Weight Mixture-of-Experts Language Model 2026 The MiniMax-M2 Series: Mini Activations Unleashing Max Real-World Intelligence 202… …
… The following papers were recommended by the Semantic Scholar API DashAttention: Differentiable and Adaptive Sparse Hierarchical Attention 2026 Self-Pruned Key-Value Attention: Learning When to Write by Predicting Future Utility 2026 MiniMax Sparse Attention 2026 Prefilling-dLLM: Predictive Prefill… …
… Inference code available at: https://github.com/kandinskylab/kvae with corresponding weights at: Audio -- https://huggingface.co/kandinskylab/KVAE-Audio Video 4x16x16 -- https://huggingface.co/kandinskylab/KVAE-3D-2.0-t4s16 4x8x8 -- https://huggingface.co/kandinskylab/KVAE-3D-2.0-t4s8 Image -- http… …
… We distill 107K trajectories from MiniMax-M2.5 and Qwen3.5-397B, introduce credit-assignment SFT to learn from productive segments of unresolved trajectories, and apply Balanced Adaptive Rollout for RL. …