DeepSeek opens its app features to developers

Inkling 2 Small uses NVFP4 for forward passes from pretraining through reinforcement learning rollouts, with higher precision for backpropagation. Compared with full MXFP8 training, the team reports an 8% wall-clock speedup in pretraining and 20% in RL on its largest jobs.

The SGLang team says NVIDIA NIM Certified microservices can use its inference engine, including the unified Radix Cache and breakable CUDA graphs.

AsyncGRPOTrainer is part of the stable API in TRL 1.18.0. Rollout generation and training run at the same time on separate GPUs, so the two stages do not have to take turns.

Ollama lists a cloud variant of Inkling 2 Small that can be called from its local client. Ollama says the service does not retain request data.

Cohere says the change brings North Micro Vision into llama.cpp and completes support for its latest architectures in the inference engine.

A summit attendee reported that SGLang serves production workloads in Meta apps, including Muse. The audience figure in the post describes the scale of Meta's apps overall.

Christopher Wolfer discusses how score centering and importance sampling can complement each other when the policy used to collect tokens differs from the policy used for training.

CEO Arav Srinivas says Perplexity handles around a billion decision model requests each month through OpenRouter. The volume is a company-reported figure.

Simon Willison released plugin support for Gemini 4 Hydrogen and compared pelican illustrations generated by three Gemini 4 models. He notes that the gap between Hydrogen and the current promotional prices for Carbon and Argon is smaller than the gap at regular prices.

Its guide covers designing, running and reporting evaluations. Recommendations include matching measures to the safety question, preregistering an evaluation plan, reporting statistics and limitations, and disclosing flaws responsibly.

Hedra post-trained Qwen3-VL-8B and tested it on 1,223 held-out clips. At a 10% false-positive rate, it caught 36% of physics glitches, compared with 20% for zero-shot use and 16% for object tracking.

The paper scales spatial compression and quantization group size together. A semi-autoregressive decoder models dependencies inside each group. The authors report about 2.5 seconds for a 1,024 × 1,024 image on one H200.

The researchers frame tasks such as segmentation and depth estimation as image editing, then fine-tune open FLUX.2 Klein models. They released the data pipeline, code and checkpoints and report 73.9 mIoU on zero-shot Cityscapes segmentation.

Machine learning points out regions a designer might want to change, and the designer can adjust local thickness during optimization. In the paper's demonstrations, linear buckling load improved by 39%, with about 15 seconds added to the design process.

A multiyear agreement with Google lets Met Office scientists test WeatherNext 3 in experimental environments. Operational forecasts continue to rely on the agency's existing physics-based systems.

The Courant center will work with academic partners on an open foundation model for the mathematical sciences. It will also study AI's mathematical foundations and model safety.

Sasha Rush and @Robertljg discuss formal checks for numerical precision, using a recently reported issue in a textbook softmax implementation as motivation.

At the time of its announcement, Tencent put Hy4 third on the overall Artificial Analysis Cyber Index and first on its CWE-Bench-AA component. The two rankings cover different sets of tasks and can change with new results.

The prototype lets a stronger model inspect a smaller model's weights in a sandbox, run a limited set of forward passes, and edit parameters directly. The author reported gains on Typed Decisions; results on other tasks remain to be established.

Riley Goodside photographed a laptop showing an AI-generated living room and inserted a real car photo into the scene. Pangram found the generated region in this example, which does not establish how reliably it will work on other composites.

Nathan Lambert discusses a new use for behavior specifications: they can help people interpret and scrutinize an agent's decisions as it takes more actions on their behalf.

Participants at the AI Engineer Code Summit published the No AI Slop manifesto. It emphasizes outcomes, understanding and human judgment over the speed or volume of generation.

The authors identify hidden actors and actions, vague wording and misplaced emphasis. They compare selected Claude Code replies with selected Stack Overflow posts; the figures describe their samples.

A developer reported that Claude Meetings treated participants outside the organization cautiously, contributing less than expected in a client call. The account describes one early-use experience.

Ethan Mollick says individual projects in his current generative AI class reached a level of completion beyond the group projects he remembers from three years earlier. This is a classroom observation.