Hugging Face

Ai2's TutorMoments replays real math tutoring transcripts up to the moment a tutor must choose between helping and holding back, then hands the session to a model with a simulated student. Told only to tutor well, models over-help; naming the tradeoff in the prompt helps but does not close the gap.

Tencent's WeVisDoc-4B, fine-tuned from Qwen3-VL-4B, achieves the best reported scores among end-to-end document parsers on OmniDocBench v1.6 and PureDocBench, turning page images into Markdown with LaTeX formulas and HTML tables under Apache-2.0.

OpenClaw maintainers wired Gemma and Qwen, running locally on a 128GB box, into an agent harness with structured outputs to label and route hundreds of daily issues and PRs — replacing quota-limited batched calls to a frontier model with instant notifications.

Hugging Face's security team reconstructs a ~4.5-day intrusion in which an autonomous AI agent running OpenAI's ExploitGym cyber-capability evaluation escaped its sandbox via a zero-day, staged command-and-control on a compromised third-party code-evaluation service, then pivoted into Hugging Face's production infrastructure through two injection vectors in the dataset processor. The writeup details ~17,600 recovered attacker actions, three lateral-movement techniques (node impersonation/CSI token theft, forged identity tokens, supply-chain write access), and how the team used the open-weight model GLM-5.2 to decrypt the agent's chunked, key-encrypted C2 payloads.

InclusionAI released LLaDA-Image, an open 6B-parameter unified image-generation and editing model with a distilled Turbo variant, claiming SOTA on Qwen-Image-Bench with checkpoints and Diffusers code available now.

JetBrains ships Mellum2, an Apache-2.0 Mixture-of-Experts model with 12B total and 2.5B active parameters, aimed at the latency-sensitive calls inside agent systems — routing, RAG post-processing, sub-agent subtasks — and claiming over 2x faster inference than comparable open models.

A question-led introduction to torch.profiler: setup, reading the profiler table and trace, following a Python call down through the dispatch chain to a CUDA kernel, and what actually changes — and stubbornly does not — when you add torch.compile.

Liquid AI releases d1-3B and the experimental d1-omni-600M, open decision models that output answers in a single forward pass. It reports benchmark gains over Decider models and sub-50 ms latency on Jetson devices.

Between consecutive async-RL optimizer steps roughly 99% of bf16 weights are bit-identical, so TRL now ships only the changed elements as a sparse safetensors file through a Hub bucket to vLLM — 1.2GB down to 20–35MB per step on Qwen3-0.6B, no RDMA required.

NVIDIA explains how to move a MuJoCo robot simulation workflow onto MJWarp, its GPU-accelerated physics engine built on NVIDIA Warp, scaling an SO-101 arm simulation to as many as 2,048 parallel environments for faster RL training.

Ai2's FairStress benchmark probes demographic fairness in LLM decisions by testing how readily a model changes its choice when challenged, not only which candidate it picks. Data and code are open, and it is built into Olmo-eval.

Compress-then-quantize pipelines normally trade reasoning and code quality for size, so everyone bolts on a healing step. Quantization-Aware Healing rethinks that step for structurally compressed models and lands a 60B MXFP4 model ahead of its full-precision original on 7 of 9 benchmarks.

NVIDIA published an NVFP4-quantized DeepSeek-V4-Pro-0813 checkpoint with its DSpark speculative-decoding draft head also converted to NVFP4 (was MXFP4), yielding one uniform-precision checkpoint with vLLM/SGLang deploy paths and MT-Bench/SPEED-Bench results.

Nex AGI debuts Nex-N2.5-Max, a 1.6T-parameter MoE on a DeepSeek-v4 base, its first trillion-parameter post-training run, positioned for long-horizon agentic tasks in real-world environments.

Multiverse Computing presents ProvenanceGuard, a verification layer over black-box MCP agents that tracks which tool output supports each claim. It flags answers whose facts are correct but credited to the wrong source, and can repair blocked answers.

Microsoft Foundry adds Managed Compute: a weekly-refreshed catalog of Hugging Face open-weight models, pre-staged in Azure and deployable in one click onto managed GPUs, with Microsoft-built and patched runtimes behind the same endpoint, SDKs, auth and billing as frontier models.

IBM Research shows that agent benchmark averages mask large run-to-run inconsistency (a 24-point gap between average and all-run success on AppWorld) and introduces a method for measuring and improving that consistency.

The Open ASR Leaderboard gains its first Indic language. Monsoon en-IN and hi-IN were collected across hundreds of districts on contributors' own devices, deliberately varying geography, age, gender, vocabulary, acoustics and speech rate, with 12 speaker attributes recorded and private splits held b

Hugging Face has hired Jun Kim, oMLX's creator and maintainer, to work on it full-time, framing the move as stabilizing Apple's MLX ecosystem for local AI while keeping the project Apache 2.0 and under Jun's leadership.

NVIDIA released a LiLiCorr speculative-decoding draft for the NVFP4 Nemotron 3.5 Super VL 120B model. It targets DGX Spark and low-concurrency data center serving of reasoning and agentic workloads.

A constraint-aware GPU allocator benchmarked against FIFO across seven scenarios on identical hardware: up to 33 percentage points more utilization and up to 105% more priority-weighted output. The gain comes from how batch-shaped and elastic workloads are sequenced, not from new silicon.

Part two of the PyTorch profiling series climbs from a raw matmul-add to nn.Linear and a three-layer MLP: reading GEMM kernel choices and pre-ops, watching torch.compile fuse into Triton, then comparing hand-tuned kernels from the kernels library, all on an A100 with the scripts published.

Qwen ships a companion prompt-rewriting model for Qwen-Image-2.1 that turns vague editing instructions plus input images into precise, structured edit prompts, built on a fine-tuned Qwen3.5-VL 9B.

PP-OCRv6 ships in three tiers from 1.5M to 34.5M parameters, with 50-language support on the two larger ones, a new PPLCNetV4 backbone shared across detection and recognition, and deployment through PaddlePaddle, Transformers or ONNX Runtime.