Hugging Face

Hcompany published Holo3.1 on June 2, 2026 as an iterative production release of Holo3 (released last March), adding quantized checkpoints (FP8, Q4 GGUF, NVFP4) built on the Qwen family, native function‑calling support,

JetBrains released Mellum2 on June 1, 2026: A 12‑billion‑parameter Mixture‑of‑Experts model trained from scratch on natural language and code that activates 2.5B parameters per token and is available under the Apache 2.

In a June 1, 2026 post by Nicholas Fuller, IBM Research argues that explicit “agent logic” — structured software primitives that steer LLMs-is necessary to run reliable, cost‑effective AI inside dynamic, long‑running

On June 1, 2026 NVIDIA released Cosmos 3, an open omni‑model that unifies world generation, physical reasoning and action generation in a single Mixture‑of‑Transformers architecture.

A beginner's guide to torch.profiler was published on May 29, 2026. Authored by Aritra Roy Gosthipaty, Sayak Paul, Sergio Paniego, Rémi Ouazan Reboul and Pedro Cuenca, the Part 1 tutorial uses a minimal matmul+add

A new step-by-step guide shows how Reachy Mini can host a full speech — to-speech pipeline on a local machine, exposing a /v1/realtime WebSocket so the robot’s conversation UI never sends audio to remote servers.

TRL now encodes only changed parameters between RL optimizer steps as sparse safetensors uploaded to a Hub bucket; vLLM fetches those deltas, shrinking per-step transfers from gigabytes to tens of megabytes and enabling

On May 19, 2026 Tom Aarsen published six Sentence Transformers CrossEncoder rerankers built on Ettin ModernBERT, from 17M to 1B parameters, and released the full training data, distillation recipe, and evaluation scripts

NVIDIA has published the Nemotron‑Labs Diffusion family: Diffusion‑capable language models at 3B, 8B and 14B parameters plus an 8B vision‑language model, with code, training recipe and a technical report.

A Dharma — AI report (May 22, 2026) on a domain‑specialized OCR model, DharmaOCR, finds that fewer than 3% of pages that never emit an end‑of‑sequence token can account for almost half of batched GPU wall‑clock time;

Dharma reported on May 22, 2026 that a 3‑billion‑parameter model, produced by a fine‑tuning pipeline, outperformed every commercial frontier API it tested on a structured OCR enterprise task while costing about fifty

Published May 19, 2026, OlmoEarth v1.1 is an updated family of transformer — based remote sensing models that lowers end-to-end compute costs by up to threefold while maintaining the original v1’s benchmark and partner —

Published May 18, 2026, PaddleOCR 3.5 lets supported OCR and document‑parsing models run on the Transformers runtime by setting engine="transformers", giving developers a direct runtime path for RAG, Document AI,

On May 18, 2026, IBM Research released the Open Agent Leaderboard and the Exgentic evaluation framework, an open benchmark that assesses entire AI agent systems — measuring performance and operational cost across

IBM’s Granite team published Granite Embedding Multilingual R2 on May 14, 2026: Two Apache‑2.

A technical post published May 14, 2026 by Rémi Ouazan Reboul, Pedro Cuenca and Aritra Roy Gosthipaty lays out an implementation for asynchronous continuous batching that disentangles CPU batch preparation from GPU

On May 11, 2026, AWS published a technical post by Keita Watanabe, Pavel Belevich and Aman Shanbhag that lays out the infrastructure and open‑source software building blocks needed to scale foundation‑model pre‑training,

A 4‑billion‑parameter model, CyberSecQwen‑4B, was published May 8, 2026 under Apache 2.0 and built for the AMD Developer Hackathon.

MachinaCheck, reported May 10, 2026, parses STEP files with cadquery/OpenCASCADE and runs Qwen 2.

A May 2026 technical preprint presents OncoAgent, an open-source, on‑premises clinical decision support stack that routes oncology queries between 9B and 27B QLoRA‑fine‑tuned models via a multi‑agent LangGraph

AllenAI announced EMO on May 8, 2026: A 14B-parameter sparse mixture — of-experts (MoE) pretrained end-to-end so modular structure emerges from data.

On May 8, 2026 Harikrishna (HK2184) published MedQA, a clinical multiple‑choice QA model fine‑tuned with LoRA on a single AMD Instinct MI300X via ROCm.

ServiceNow engineers found that migrating PipelineRL inference from vLLM 0.8.5 to a vLLM V1 rewrite (vLLM 0.18.1) produced rollout logprobs the trainer did not expect, shifting early GSPO metrics.

On May 6, 2026, Open ASR Leaderboard maintainers said several high‑quality English test sets from Appen Inc. and DataoceanAI will be hosted privately to reduce test‑set contamination and “benchmaxxing.