Hugging Face

IBM Research's case that frontier context windows are the wrong answer to dynamic, policy-constrained enterprise workflows: put software primitives — knowledge graphs, program analysis libraries, algorithms — in the agent harness to steer the model and shrink its context, illustrated across four IBM

Cohere Labs released Tiny Aya Base 32K, a compact multilingual model covering 46 languages with a 32K context window, aimed at researchers and low-resource-language builders who need long-context multilingual generation without a large-model footprint.

vLLM's transformers modeling backend now matches or beats vLLM's hand-written implementations on Qwen3 4B, 32B and a 235B FP8 MoE, behind one --model-impl transformers flag that composes with tensor, data and expert parallelism. Linear-attention models are not yet supported.

Google released GNM v3, an Apache-2.0 parametric 3D head model trained on high-resolution scans, giving disentangled control over identity, expression, and pose along with articulated eyes, teeth, and tongue, with native NumPy, JAX, PyTorch, and TensorFlow support.

DPO used as failure-mode surgery rather than alignment: DharmaOCR paired correct transcriptions against the model's own degeneration loops after SFT and cut degeneration in every VLM family tested — 59.4% average, 87.6% best case — because sequence-level preference penalizes what token-level loss ca

NVIDIA published a DFlash 2 draft checkpoint for the Muse Spark Open 816B-A42B multimodal MoE model. It adds in-block convolution and a pairwise path selector to speculative decoding, with serving instructions for vLLM and SGLang.

Hugging Face's account of reproducing 2,200 ICML papers, with per-project logs, code and traces published in an interactive logbook.

Microsoft Research released VibeVoice-ASR-Streaming in a 1.5B configuration (3B params). It transcribes who said what as audio arrives, accepts customized hotwords for names and technical terms, covers ten languages, and ships MIT licensed with a Transformers pipeline.

Part three of the PyTorch profiling series takes attention apart under the profiler: naive scores-mask-softmax, in-place causal masking, then SDPA's math, efficient, flash and cuDNN backends, with the scripts and A100 traces for each so you can see how the kernels differ.

Hugging Face Transformers can now load and run GGUF quantized checkpoints directly through from_pretrained, reusing llama.cpp's ggml kernels for near-native performance on Apple Silicon, with benchmarked size/precision tradeoffs.

Microsoft Research released VibeVoice-ASR-Streaming in a 7B configuration (9B params), the larger of the two sizes. It streams who-said-what transcription, accepts customized hotwords for names and jargon, covers ten languages, and ships MIT licensed with published code.

Tencent released Hy-Nomas, an Apache-2.0 segment-autoregressive diffusion transformer that generates video worlds turn by turn from text and camera trajectories. It outputs 720p at 24 FPS with 8 denoising steps and has an INT8 single-4090 mode.

Ai2 details a GPU cluster scheduler built on team budgets, fair-share queue ordering and a declared-runtime contract. It covers the incentive failure of the old priority system, rollout lessons, and a 30-day test showing near-full occupancy.

Browser caches are origin-scoped, so two sites using the same Whisper checkpoint each pay 177MB and each ship their own copy of the ONNX Wasm runtime. This walks through the proposed Cross-Origin Storage API in Transformers.js: hash-addressed blobs, integrity by design, permissioned reads.

NVIDIA releases Kumo Tabular, an open tabular foundation model in three sizes. It does in-context classification and regression in a single forward pass, and it claims top rankings on TabArena, BeyondArena, TALENT and ScoringBench.

Treble and Hugging Face open a far-field ASR leaderboard built on wave-based simulation of 14 rooms validated against real measurements, with held-out audio and standardized evaluation hardware. First results: far-field WER at low SNR is several times near-field WER on the same speech.

inclusionAI releases Ling-3.1-flash, an open hybrid-linear MoE reasoning model with 560B parameters, 25B active, and 1M context. It targets agentic coding and tool use and ships FP8 weights and a LingCode harness.

Tencent's Simple Attention Sparsification trains a differentiable gate router to select KV blocks end-to-end with the LM loss, and ships ready SGLang-compatible gate checkpoints for Qwen3-4B/8B/14B for long-context sparse decoding.

Liquid AI shipped 4-bit checkpoints trained by distilling a high-precision teacher into a quantized student, rather than quantizing after the fact. The result: Q4_0 speed and size at Q4_K_M-or-better quality across four model sizes, profiled on real edge hardware.

A working glossary for the agent terms nobody defines the same way: model, scaffolding, harness, context engineering, policy, tool use, skills, sub-agents, plus the RL-side vocabulary of environment, trainer, rollout and reward — with the scaffold/harness distinction pinned to where it matters.

Amazon's ALoDLM-8B is a diffusion language model built on a Qwen3 backbone that refines unresolved tokens through adaptive recurrent passes and commits confident tokens in parallel. The card reports benchmark results and B200 throughput, with code released and a non-commercial license.

Multiverse Computing's paper shows topic-level safety guards can't separate benign from harmful prompts within one topic, and trains on prompt pairs sharing a topic anchor but differing in intent to sharpen where a model refuses versus answers near that boundary.

Ant Group releases Ling-3.0-flash-Fin, a 124B/5.1B-active MoE model tuned on financial data for source-grounded research, multi-document reasoning, and spreadsheet workflows, evaluated across FinFIRST, FinCRAFT and other finance-agent benchmarks.