Open-weights multimodal LLM checkpoint designed for 1M-token long-context agent runs that combines large-scale on-policy RL with groupwise grading for iterative self-improvement. Key traits: sparse MoE backbone (≈1.02T total / 42B active), text-only output with image/video/audio inputs, MIT-licensed weights on HuggingFace—suited for long-horizon agents and research at significant infra cost.
A 9B agentic multimodal SFT checkpoint distilled from Qwen3.5-9B for coding, general agent tasks, visual coding and cybersecurity. Provided by Xiaomi MiMo as a research seed (77.4B-token SFT mix) to bootstrap agentic RL and tool-use experiments.
A multimodal supervised fine‑tuning corpus for training MLLMs: 44.3M samples (100.3M QA pairs) over 19.7M image groups, annotated with short/long chain‑of‑thought and produced by rerunning the HoneyPipe curation pipeline with multi‑model rollouts and verifications.
A distilled LoRA adapter for Qwen-Image-2.1 that runs text-to-image generation and instruction-driven image editing in a few transformer passes (shipped as a 6-step r256 LoRA). Samples with a fixed sigma schedule, no classifier-free guidance; non-commercial research license.
Uses diffusion-model forking moments as a proxy for perceptual distance to automatically generate pointwise reference-grounded labels, enabling annotation-free training of reference-based image quality assessment metrics.
Performs speaker diarization (who spoke when) for live and recorded audio using an open-weight, 100M-parameter streaming-capable model that supports up to eight anonymous speaker channels, overlapping speech, chunked processing, and configurable latency for ASR integration.
Provides 2,000 synthetic multiple-choice items designed for continuation log-likelihood scoring to evaluate small language models' Theory of Mind (social-cognitive) abilities; 40 constructs, balanced answer positions, and easy/medium difficulty.
Turns a context, a question, and 2–20 candidate answers into a single chosen option for classification, routing, ordered scores, and Boolean decisions. Built on mmBERT-small with a 144.3M-parameter decision head, runs on CPU with up to 8,192 combined tokens; FP32 weights occupy 550.5 MiB and are Apache‑2.0 licensed.
Performs schema-driven, low-latency classification and structured decision-making over English text. Supports multi-head scoring, constrained joint decoding with confidence/feasibility metadata, span extraction, and local CPU/GPU deployment via the gliner2 runtime.
Enables native joint video-and-audio generation and training up to 2K using a dynamic block-sparse attention scheme that cuts attention cost while preserving audio–visual coupling; provides preview checkpoints, training and inference code. High-resolution training requires large multi-GPU (80GB) clusters and CUDA ≥ 12.4.
Provides 5.5K+ self-contained data-analysis RL tasks: each row bundles a real tabular dataset, a question, and a deterministically-gradable gold answer. Verified from jupyter-agent notebooks; splits for training, held-out testing, and quick eval; intended for prompting, fine-tuning, and agent RL.
Provides imagined interaction segments generated by world models for RoboTwin2.0 tasks, stored as fixed-length HDF5 chunks (21 observation frames, 20 actions, rewards and episode flags). Useful for training and evaluating world-model-based policies; currently limited to the RoboTwin2.0 subdataset.