From the archive
AI models are the systems behind chatbots, coding assistants and tools that work with text, images, audio and video. Follow new language model releases, open-weight models, local AI and benchmark results. Our coverage looks at capabilities, context limits, access, pricing and evaluation methods so you can compare the evidence behind a launch, not just the claims in its announcement.
Lead story
OpenAI’s Decisions API is described as a limited-preview service for asking Luna to choose among predefined answers, while a new Hugging Face community guide offers developers a practical way to think through building decision workflows. The key distinction: the guide says the announcement does not establish a public…
Read the report & discussion ↗Moonshot AI/Kimi Watch
Moonshot AI says its new open-source Kimi K2.6 model can coordinate agent swarms of up to 300 sub-agents across 4,000 steps at once, a substantial increase on K2.5. It is also pitching longer-running coding work as a strength, with examples that put the model through thousands of tool…
Geordie Watch
Wood Partners is using Geordie to inventory AI agents across multiple platforms and enforce policies against risks classified as critical, according to a new customer case study. The practical shift is from separate platform dashboards to a shared picture of agents, their connections and the controls applied to…
NVIDIA Watch
NVIDIA is adding a 64GB unified-memory option to its DGX Spark AI workstation, with sales through partner manufacturers due to begin on 23 October. The useful twist is that two systems can be linked to pool 128GB of memory, giving local AI developers room to try larger workloads…
LocalAI Watch
LocalAI 4.11.0 brings several new capabilities to the local AI project, including speaker-labelled transcription, sound detection alongside speech, a Decisions API, model failover and text-to-animation with Kimodo. For people running AI workloads themselves, this is a release with several things to try, not just a version number doing…
Watch Desk
A Hugging Face community guide compares six reinforcement-learning methods used in language-model training, matching each to the problem it is intended to address. Its useful point: PPO, RLOO, GRPO, Pref-GRPO, GSPO and GDPO are not interchangeable recipe names; they make different trade-offs in compute, training stability and reward…
Watch Desk
Braze has launched several AI tools for marketing teams and signed a strategic collaboration agreement with Amazon Web Services. The package combines campaign decisioning, automated checks and conversational agents, with a new route for bringing Braze functions into other AI tools.
Watch Desk
A new class of AI models is being built to handle the quick, structured decisions around larger language models: routing requests, checking outputs and choosing what an agent should do next. The potential payoff is faster, cheaper routine checks, with a more capable model called in when the…
OpenAI Watch
OpenAI’s “Sign in with ChatGPT” lets users of participating third-party apps authorise AI requests against their existing Plus or Pro usage allowances. The change could make it easier to try compatible tools without setting up separate API keys, while leaving users to manage how much of their plan…
DeepSeek Watch
DeepSeek has released preview desktop apps for DeepSeek Harness, its open-source agent framework, with installers for Apple silicon Macs and 64-bit Windows PCs. The practical shift is that users can now try the harness through a desktop interface, with tools and plugin management built in, rather than assembling…
Watch Desk
Fine-tuning improved one narrow tool-use task in SCAI’s five-week experiment, but made a strong model worse at coding. The lab’s practical conclusion is refreshingly unspectacular: measure the problem, try cheaper fixes, and fine-tune only when you can show they have run out of road.
DeepSeek Watch
DeepSeek and Huawei have launched an open-source programming toolkit aimed at Huawei’s Ascend AI chips, offering developers an alternative software stack to Nvidia’s CUDA. The toolkit includes TileLang plus compute and communication libraries, according to Mobile World Live’s 30 September report.
Microsoft AI Watch
Microsoft’s ThinkingBox benchmark grades AI agents on the state they leave behind, then repeats each task 20 times to see whether success holds up. Across 507 business workflows, the results suggest that a fluent final answer and a successful tool call are poor substitutes for checking the records.
Google DeepMind Watch
Google DeepMind has released Gemini Robotics 2, a robotics model suite combining an embodied reasoning model with on-device controllers. Research lead Keerthana Gopalakrishnan says the work also tackles whole-body control, while the field still has a long way to go on getting systems to generalise across different robot…
