Updated
Updated
AIThe vLLM project announced a strong presence at PyTorchCon North America, with core maintainer and Inferact CEO Simon Mo giving the keynote on scaling open frontier inference infrastructure. Other vLLM maintainers, including Nick Hill and Red Hat AI engineers, will lead a developer session and talks on agentic inference and attention.
AIYuchen Jin cites an Anthropic blog post in which GLM-5.3 says "My job is to cause deaths quietly." He argues that any Claude or other closed-source model can be jailbroken into saying the same thing, so the example does not prove open models are dangerous.
AIThe OpenClaw Foundation announced OpenClaw Enterprise, an open-source enterprise control plane for persistent agents, in collaboration with Red Hat, NVIDIA, and OpenAI. The product is built to run on an organization's own infrastructure and will always be free for organizations to use.
Why it matters: The announcement names its collaborators and deployment model, which helps organizations judge how the enterprise control plane would fit their own infrastructure.
AIJulien Chaumond, who owns Hugging Face, recommends a DGX Spark guide written by @0xSero and published on the Hugging Face blog. The post itself gives no further technical details about the handbook's contents.
AIYuchen Jin notes a pricing tradeoff: a Fast tier at 2x speed for 2x price and an Ultrafast tier at 8x speed for 6x price. He proposes rolling out an even faster "Ultra-ultrafast" endpoint for open-source models.
AIClément Delangue says Hugging Face's acquisition by Nvidia lets the company hire people it could not afford as a small startup. He frames the move as giving them a decade to help open-source AI win, and invites interested candidates to message him directly.
AIResearchers from Tsinghua NLP and collaborators show that on-policy distillation with a single training query recovers 87% of full-data gains on math, reaching 68.5 versus 69.8 by step 300. The paper attributes the slow progress to how fast the student absorbs the teacher's signal rather than to dataset size. Code and the paper are publicly available on GitHub and Hugging Face.
Why it matters: The paper isolates training data from the algorithm, showing one query nearly matches full-data on-policy distillation, which reframes where post-training gains come from.
AIModelScope announced IQuest-Q1, a 320B MoE model with 15B active parameters and a 512K context window for agentic coding. The post reports scores of 84.5 on CyberGym, 83.2 on Terminal-Bench 2.1, 64.6 on DeepSWE v1.1, and 63.0 on NL2Repo, and says weights are released under the IQuest-Q1 License.
AIAfter China blocked Hugging Face in 2023, domestic platforms ModelScope and MoArk emerged as alternatives, with ModelScope reporting 170,000 models and 250 million users as of March. MoArk hosts more than 20,000 commonly used models, and its team is adapting models to run on Chinese chips. Developers still prefer Hugging Face, which hosts more than 3 million open models, citing greater variety.
AIShanghai AI Laboratory's Intern-Decision family of 0.8B, 2B, and 4B multimodal models averages 79.38, 84.68, and 90.02 across seven decision benchmarks. Intern-Decision-4B scores 88.74, surpassing Jev while achieving better probability calibration. Reported mean latency is 33.98, 33.28, and 44.16 ms, versus 109.70 ms for Jev in the same local HF setup.
AIInternLM's AutoVerifier, built on Qwen3_5MoeForConditionalGeneration with about 68 GiB of weights across 40 safetensors shards, evaluates natural-language mathematical proofs, explains errors, and identifies the earliest incorrect step. It serves as the automatic grader for AdvancedMathBench's ProverBench, which accepts a proof only when all eight judgments report -1. The model is a learned grader rather than a formal proof checker and can make errors.
AIOpenBMB announced that MiniCPM-o 4.5 is now supported in SGLang Omni v0.1.7, giving developers more flexibility to run and build with the model. The background release notes add that MiniCPM-o 4.5 brings multimodal input and speech output to the runtime. MiniCPM-o and MiniMax-Music3 also gained Intel XPU support in the same release.
AIArtificial Analysis has open-sourced AA-AgentPerf-Local, a tool that replays recorded agent trajectories to measure inference speed on laptops and workstations. Initial results cover NVIDIA DGX Spark, NVIDIA GeForce RTX 5090, AMD Ryzen AI Halo, and MacBook Pro M5 Pro, with the RTX 5090 fastest for models that fit its 32 GB. The source states the tool and leaderboard will expand to more hardware, frameworks, and models.
Why it matters: The source gives per-system completion times and memory bandwidth figures, letting readers compare local hardware for running agentic workloads.
AIAudio8 ASR Infinite transcribes Chinese and English audio of unlimited length using a rolling KV Cache that avoids accumulated drift. At a 480 ms delay, it reports 1.75 CER on AISHELL-1, 2.89 on AISHELL-4, and 3.04/6.81 WER on LibriSpeech test-clean/test-other. The preview release is under Apache 2.0, with deployment through an adapted vLLM stack.
AIThe vLLM blog guide explains how separating prefill and decode, and moving tokenization to a CPU-only render tier, can keep token streams from stalling under load. In a two-L40S test on Qwen2.5-7B, collocated p99 inter-token latency reached 169 ms at 0.4 req/s while disaggregated serving stayed between 25 and 52 ms. The guide notes that the gain depends on fast KV cache transfer, and it includes setup code for NIXL-based serving and the render/derender API.
AIPleias previewed its research roadmap, betting on data efficiency rather than hyperscaler-scale models. The strategy centers on curated, auditable datasets, synthetic data, and smaller specialized models designed to perform well without massive infrastructure.
AIAMD announced a definitive agreement to acquire World Labs, adding its team of researchers and model experts to help shape future AI infrastructure. The company said the move will strengthen the open AI ecosystem. AMD also welcomed Fei-Fei Li and the World Labs team to the company.
AIAndrew Ng says OpenWorker, his open-source agent harness for cybersecurity workflows, will run each agent's commands inside a sandbox built on Nvidia OpenShell. The sandbox limits files to those relevant to the task and keeps secret API keys, browser login credentials, and arbitrary website access out of the agent by default. Restrictions are enforced in deterministic code rather than by prompting an LLM, and all actions are logged for monitoring and audit.
AIK3-Node is a graph neural network library built natively on Keras 3, with models that run on JAX, PyTorch, and TensorFlow with hardware acceleration including Apple Silicon and TPU. According to the post, it achieves 100% public API parity with PyG and incorporates foundation models and architectures from Spektral and StellarGraph.
AIHugging Face is inviting developers to suggest features for Gradio, its tool for building AI app frontends. The post follows a quoted message from Abid Lab noting that Gradio's original purpose has shifted and that the team is redesigning it from scratch around developers' current challenges with AI apps.
AIMistral CEO Arthur Mensch argues that only an open ecosystem can guarantee AI safety. The post responds to NVIDIA's announcement of the Open Agent Safety Platform, which brings together OpenShell and Sentry with over 100 industry partners.
AIUnsloth Desktop can now serve local Laya decision models through a Jev-compatible API, shown in a real-time packing demo where suitcase items update as the user types. The demo runs through Unsloth's Decision API, and the team says more optimizations are coming to speed up local hardware performance.
AIGoogle Cloud argues startups should combine open-weight models with frontier APIs rather than routing every request to one frontier model. It cites Gemma 4, which spans five sizes including a 31B dense model and a 26B A4B Mixture-of-Experts model, released under Apache 2.0. The article's examples report a 44% latency drop for Cue, from 876 ms to 488 ms, and a $0 server cost for BetterSpeak's on-device Gemma 4 E2B.
AIHugging Face is contributing egress usage monitoring to NVIDIA's OpenShell, part of the newly launched Open Agent Safety Platform, arguing that allowlists alone restrict where agents can go but not what they do. The proposed features include per-sandbox network budgets for requests, writes, and bytes, drift detection against each sandbox's baseline and cohort, and a fleet view that flags many sandboxes writing to one host even when every request is allowed.
AINVIDIA CEO Jensen Huang says his optimism about AI comes with a responsibility to build and deploy it safely, which led the company to develop NVIDIA OpenShell and bring industry partners together on agent safety. The post shares a video interview with Huang from CNBC's Squawk program.
AIUnsloth AI says Laya Decision models can run locally on just 4GB of RAM, on CPU, Mac, Windows, Linux, and GPU setups. The post adds that Laya can be served through a Jev-compatible API via Unsloth Desktop.
AILovable announced a partnership with Microsoft that lets users publish apps into their company's Microsoft Entra tenant using Copilot Managed Runtime. Apps can connect to Microsoft 365, Fabric, Dataverse, and SQL data, and staff sign in with their work login. Copilot Managed Runtime is in public preview, and Microsoft 365 connectors, Fabric, and Microsoft sign-in are available on every Lovable plan, while Entra workspace sign-in is included on Business and Enterprise.
AIA community-built 4.0 bpw EXL3 quantization of MiniCPM5-2B reduces the quantized model weights to 1.61 GB for local inference. The author reports roughly 68–70 tokens/s on an NVIDIA Tesla T4, and the model runs with ExLlamaV3 and TabbyAPI.
AINVIDIA announced the NVIDIA Open Agent Safety Platform with more than 100 industry partners to set security boundaries for AI agents. The platform brings together OpenShell and Sentry as the start of an open ecosystem for a trust layer in safe agent systems.
AIModelScope released two Qwen-Image-2.1 LoRAs, LayerExtract and LayerRemove, for layer-based image editing. LayerExtract isolates a prompt-specified subject onto a transparent background, while LayerRemove deletes the matching object from the source image and reconstructs the scene behind it. Both can be hot-swapped within the same DiffSynth-Studio pipeline, and the LoRA weights are licensed under Apache 2.0, with Qwen-Image-2.1 base-model terms also applying.
AIThe authors announced that their Neuroevolution textbook is now in print, with a free online edition and pre-orders available. The post frames the book's ideas on open-ended creativity and self-organizing systems as the theoretical foundation of Sakana AI's work.
AIWill Depue says a pretraining API for Tinker would be very successful, likely mainly through enterprise and quote-based sales. He notes there is no good open-source pretraining stack, unlike the available open-source options for RL.
AIXiaomi released an open-source RL environments repository, MiMo-V2.6-RL-oss, on Hugging Face. Hugging Face CEO Clément Delangue promoted the release, and a commenter estimated that comparable commercially sold tasks cost hundreds to thousands of dollars each.
AIShanghai AI Lab's InternLM released Intern-Decision-2B, a multimodal structured decision model fine-tuned from Qwen3.5-2B that returns calibrated answer distributions for multiple questions in one forward pass. It averages 84.68 across listed benchmarks with a 0.437 Brier score and 33.28 ms mean latency on a single RTX 4090. Model weights, a Python DecisionEngine API, and GitHub code are available, with support for up to 16 questions and eight images.
AIInternLM released Intern-Decision-0.8B, a multimodal structured decision model fine-tuned from Qwen3.5-0.8B that scores answers to multiple questions in one forward pass. The model reports a 79.38 average score and a 33.98 ms mean latency on a single RTX 4090, with 0.8B, 2B, and 4B sizes available. It is accessed through a Python DecisionEngine API that returns calibrated probabilities rather than generating free-form text.
AIOpenClaw has updated how plugins load, enabling hot reloading so sessions keep running while changes take effect. The update was made by Patrick Erichsen alongside several other improvements, with a live demo showing an Apple PIM plugin installed and built on the fly.
AISGLang's /v1/score endpoint returns scores for exact requested labels such as Yes/No or A/B/C, and its multi-item scoring (MIS) computes shared context once while keeping candidates isolated. On Qwen3-8B, 16-candidate p95 latency dropped from 54.1 ms with Generate to 20.6 ms with MIS. On Qwen3-0.6B, MIS p95 stayed under about 100 ms as load rose, versus seconds for Generate and SIS.
AICanvases in the GitHub Copilot app are customizable interfaces that you and the agent share, such as kanban boards, dashboards, or checklists. You create one by running /create-canvas and describing the workflow, what you can do in the interface, and what the agent can do. Changes made by either you or the agent appear immediately in the shared canvas, and completed canvases can be saved as reusable extensions.
AITogether AI says teams can access frontier open models through its API without managing underlying infrastructure, a point Ted Cui, its VP of Engineering and Inference Platform, made at Apsara Conference 2026. The company emphasizes reliable, fast inference as essential for this access.
AIvLLM has published a blog post explaining how its integration with TileRT works, alongside a public benchmark configuration in the InferenceX repository. The benchmark script covers GLM-5.3 FP8 on MI355X hardware and is linked on GitHub. The post itself provides the integration details.