AI Model2026
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.
AI Audio2026
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.
AI Model2026
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.
Shows that Transformer LLMs output a superposition of next-token distributions when token embeddings from distinct text streams are linearly combined, finds this property is architectural but weakened by pretraining, and demonstrates it can be restored by lightweight fine-tuning and disentangled at decode time.
Trains decoders and diffusion generators to be robust to random subsets of pretrained encoder layers by regularizing layer-fusion during training, narrowing the reconstruction–generation gap in representation autoencoders and improving ImageNet-256 PSNR and gFID.
Reduces self-attention complexity to O(N log N) by using a coarse-to-fine (pyramid) Top-K block selection with LogSumExp scoring, implemented with hardware-aware Triton kernels for fused routing and scoring—aimed at long-context LMs and retrieval tasks.
AI Model2026
Open-weight 309B Mixture-of-Experts causal LLM with 15.5B active parameters and a native 1M-token context for coding and AI R&D. Combines Sliding-Window Attention and DeepSeek Sparse Attention (no full-attention layers), supports FP8 inference; weights under MIT license.
Improves test-time scaling of looped transformers by adaptively assigning extra recurrent iterations to tokens that benefit most. TaH2 is a post-training method that jointly trains an iteration decider with the backbone using lookahead depth supervision, boosting the accuracy–compute slope and peak accuracy on challenging benchmarks.
Demonstrates that pretrained transformers typically use only ~1–3 lines of depth to follow reference chains, and that a task‑trained rank‑8 LoRA applied at one early layer (with all other weights frozen) can extend reference‑following to dozens or hundreds of lines while adding only a few ten‑thousand parameters.
Turns solved protein structures into FoldingCorpus and Fold2Reason — a post-training recipe that supervises an LLM with discrete structural Q&A plus continuous 3D geometry to improve spatial, graph and scientific reasoning; reports consistent gains across 10 benchmarks.
AI Model2026
Delivers fast, calibrated 'System 1' decisions over text and images (yes/no, 2–256 choices, 0–5 scores) in one forward pass, plus full 'System 2' Qwen3.8-27B multimodal reasoning. Highlights: zero-shot vision judging, 256K-token context, vLLM decision API, Apache-2.0.
AI Model2026
Parses documents and answers questions about their contents with a 0.8B-parameter model. It combines OCR, structured extraction, and visual question answering, including support for scanned, skewed, and photographed pages.