Hugging Face

NVIDIA released Nemotron 3 Diarization, an open-weight 100M-parameter speaker-attribution model ranked #1 on VoiceArena's benchmark, supporting up to eight speakers in real-time streaming or offline modes.

Hugging Face and Google's TorchTPU gives PyTorch a native TPU backend. Transformers, Diffusers and Kernels run on the tpu device without torch_xla's lazy tensors or a JAX port.

DeepSeek published the V4.1-Flash-1017 checkpoint on Hugging Face under the MIT license. It has fp8 weights, an image-text-to-text pipeline tag and vLLM and SGLang serving instructions. No benchmarks are on the page.

Hugging Face launches an Open TTS Leaderboard scoring multilingual text-to-speech and voice cloning with objective metrics: ASR-based error rates, batched and streaming latency on H200 and CPU, and speaker similarity. It complements human-preference arenas rather than replacing them.

Hugging Face rebuilds most of AUTOMATIC1111 as 'Workflow1111,' a 73-node Gradio Workflow canvas covering txt2img, hi-res fix, img2img, ControlNet-style annotators, inpainting, and img2video, each node an API-exposed fn/model/Space/dataset call.

Reframes LLM depth pruning as an Ising optimization problem solved with physics-inspired solvers, beating the best existing block-removal method by nearly 23 MMLU points at 50% compression of Llama-3.3-70B.

LeRobot v0.6.0 adds world-model policies that learn to imagine future frames, several new VLAs, a reward-model API, six simulation benchmarks behind one eval CLI, a rollout CLI with DAgger-style human corrections, FSDP training and depth support.

A walkthrough of AWS's Apache-2.0 Strands Robots SDK: record demonstrations in simulation as LeRobot datasets, push to the Hub, run a policy, then move the same agent code to a physical SO-101 by changing one keyword argument, with a Zenoh mesh for fleets.

IBM Research sets ALTK-Evolve against ACE. Both turn past agent trajectories into reusable lessons with no weight updates, and both refuse to compress — brevity bias and context collapse destroy the detail. They split on delivery: one growing playbook versus retrievable, cluster-merged guidelines. T

H company released NeoMME, 260M and 800M multilingual encoders that run image patches and text through a single bidirectional Transformer trained from scratch, with no vision tower or pretrained decoder. The retriever variant returns dense and late-interaction embeddings in one pass under Apache 2.0

Reflection's Beam is an open 501B mixture-of-experts model (23B active, 1M context, Apache 2.0) built for coding and agents. Hugging Face covers its architecture, BF16/FP8/NVFP4 checkpoints, inference support, and SFT and RL post-training recipes.

ServiceNow describes AutoSynthData, which turns a target model's failures and a teacher model's successes into new agentic tasks with verifiers. It adds sample-level repair and batch review, and is demonstrated on the EnterpriseOps Gym benchmark.

MiniMax published M3.1 Flash weights on Hugging Face: a multimodal mixture-of-experts model tagged for agents, coding and video, shipped in 8-bit MXFP8 with Transformers and vLLM serving instructions under a community license.

IBM's build notes for Granite 4.2: dense 3B/8B/30B reasoners pretrained from scratch on ~15T tokens out to 512K context, SFT on chain-of-thought and agentic trajectories, then staged RL where the 8B and 30B learn tool use in real sandboxes. Apache 2.0, with OpenCode and OpenHands recipes.

IBM Research on why model routing stops being a classification problem in production: agent workloads reuse enormous context, so cache-read rates flipped the cheaper model on 417 AppWorld tasks; task difficulty is unknowable at routing time; and latency and compliance constraints all bind at once.

Hugging Face's PyTorch team details MPS improvements in PyTorch 2.15: FP8 on Mac, faster linear and softmax decoding paths, and new Metal kernels. MoE decoding reportedly jumps from 9.4 to 61.3 tokens/s on Qwen3.6-35B-A3B.

UK AI Security Institute is adopting EvalEval Coalition's shared evaluation-reporting infrastructure (the Every Eval Ever schema), aiming to make government model evaluation results reproducible and verifiable rather than siloed one-off reports.

NVIDIA's Magpie multilingual TTS is a 364M-parameter open-weights model spanning 12 languages — Arabic, Korean and Brazilian Portuguese are new — with male and female voices per language and NIM containers for self-hosted, low-latency speech inside cascaded voice agents.

IBM's ALTK-Evolve distills a playbook from an agent's training-task attempts and adds it to the system prompt. With GPT-4.1 on a stock ReAct scaffold it reports 67.5% Pass^4 on tau2-bench retail, above the leaderboard entries, with no weight updates.

An open reproduction of Surya Narreddi's watercolour-painting model, built with TRL and OpenEnv. Covers building the RL environment, using a hand-rated reference pool as the reward, comparing three reward mixes, the infrastructure that broke, and what the training run cost.

Every Gradio Space now publishes an agents.md listing its schema, call and poll endpoints, file upload and auth — enough for an agent to drive it without a client library. The author chained an image Space into a 3D reconstruction Space to build a gallery from two prompts.

InclusionAI released Ming-Image-0.1-Design, a 6B text-to-image model tuned for posters, infographics and UI mockups, with native RGBA transparent-background generation and recommended vLLM-Omni deployment settings.

Hume's Real World VoiceEQ scores 40+ proprietary and open voice models across ASR, TTS, speech-to-speech and understanding using over a million human ratings, and finds models speak far better than they listen while saturated benchmarks overstate real-world quality.

OlmoEarth Studio can now compute and export embedding vectors for any polygon and time range, delivered as int8 Cloud-Optimized GeoTIFFs from Nano, Tiny or Base encoders over Sentinel-1 and Sentinel-2 imagery. Ai2 documents downstream uses, a fine-tuning path, and the limits.