Research Papers Kimi K3 is an open Mixture-of-Experts model with 2.8T parameters (104B activated per token), native vision input and a 1-million-token context window, built on Kimi Delta Attention and post-trained with reinforcement learning across general, agentic and coding tasks.
Xiaomi's robotics team presents a vision-language-action foundation model for mobile manipulation, pretrained on more than 100K hours of real-world UMI trajectories with auto-labeled scene-transition language, then post-trained to match embodiments and human imperative prompts. The paper claims clean scaling with more data and parameters, out-of-the-box performance in unseen environments and efficient fine-tuning for dexterous tasks.
Symbolic music models plan composition explicitly but never sound like a finished recording, while audio models produce full songs that keep the composition hidden. YuE2 combines both in a single AR-NAR Mixture-of-Transformers that plans symbolically, expands the plan into semantic music tokens, and renders the full song, using two new representation models (MERT2, SheetSage2) to learn from recordings that lack aligned scores. The authors report a SongBench Global Avg of 6.73 on WildSongBench (6.96 at best-of-8), and in expert listening the best-of-8 samples were preferred over Suno v4.5 and roughly tied with Suno v5; the readable score also makes the output editable by external language-model agents.
VoiceMem gives duplex speech models a working memory that keeps facts and emotional context across a conversation, so a voice assistant can remember both what you said and how you said it, in real time.
NeoHorse-1 records, for every user turn, the predicted capability demand, the selected service tier and how the interaction went, then converts these records into training examples admitted through structural validation and multi-dimensional semantic checks, feeding them back as the next round of agentic post-training.
MonkeyOCRv2 pretrains on MonkeyDoc v2, a 113-million-image document corpus spanning 17 languages, combining image-to-text generation with pixel-level stroke reconstruction, so the visual encoder acquires the character-level perception that document tasks demand.
Frontis-MA1 ships with OpenMLE, an open full-stack system (task gyms, RL operator learning, evolutionary search) for studying recursive self-improvement. Its 35B meta-evolution agent, post-trained around Draft, Improve, Debug and Crossover program-evolution operators, claims MLE-Bench Lite results the authors say exceed GPT-5.5 plus Codex, with components transferring to held-out NatureBench Lite.
Edge0 is an open-source inference engine that serves a 35B mixture-of-experts model from a single 24GB GPU by loading expert weights from SSD on demand; a trained per-layer prerouter predicts the next layer's expert choices one token ahead so SSD reads overlap with compute, and a recovery LoRA pays back the quality lost to int4 quantization. The authors report 20 tok/s at about 3GiB of peak active memory, within a few points of the fp16 teacher on five public benchmarks.
Streaming video assistants must watch a live feed in real time and still remember what happened minutes ago. OneStreamer (Nanjing University) trains a 4B model to do both with one shared generation process: while watching, it proactively writes time-stamped captions and summaries into a hierarchical text memory, so follow-up questions are answered from those written records instead of re-encoding old frames. It also releases OneStreamer-1M, a one-million-record dataset for streaming video interaction, and reports best results among compared methods on all eight evaluated streaming understanding benchmarks.
Most VLA pipelines route visual observations through a large language model before decoding actions, which is heavy at inference time. TurboVLA instead encodes vision and language independently, exchanges information through a lightweight bidirectional interaction module, and decodes continuous action chunks directly — a 0.2B-parameter policy the authors report reaching 97.7% average success on LIBERO in 31.2 ms with under 1 GB of VRAM on an RTX 4090.
Real-time robot control needs enough visual history to judge motion and task progress, but processing that history usually slows the action. Long-WAM, NVIDIA's entry in the LongLive family, shows the payoff depends on how the video foundation was pretrained: with autoregressive pretraining, stretching context to about 19 seconds lifts RoboCasa success rates clearly, while a bidirectional init gains nothing. Streaming observation encoding, asynchronous execution and hardware-specific acceleration keep it on time budget on RTX 5090, DGX Spark and Jetson AGX Thor, with real-robot demos on Unitree G1 and YAM.
AuK unifies speech generation and editing in a single open-source model: you describe the desired change in natural language, and it either synthesizes new speech or modifies existing audio accordingly.
Reinforcement learning on diffusion models only rewards the final image, leaving intermediate denoising steps unsupervised. This framework converts the endpoint reward into explicit bounded targets for each sampled query state, training the model on-policy through self-distillation.
Automatic transcription has long worked on solo piano or isolated tracks but falls apart on real, full-band recordings. MuScriptor (Meta) targets exactly that setting: its recipe pre-trains on synthetic audio, fine-tunes on real music, and adds reinforcement-learning post-training, plus a conditioning mechanism so users can specify which instruments should appear in the output. The authors release the model weights openly as an ISMIR 2026 camera-ready.
OpenWAM breaks the tightly coupled world-action model into swappable modules with unified training, deployment and evaluation, turning pretraining into a controlled experimental program where each design choice can be isolated and tested.