Open weights21 Sep 2026
Alibaba open-sources Qwen-Image-2.1, a 7B image model topping open leaderboards
Qwen-Image-2.1 unifies text-to-image generation and editing in a 7B visual generator with native RGBA transparency and up to ten reference images, scoring 60.28 on public benchmarks and edging out larger closed systems.
Our takeCompact open models raise the bar for image datasets: cleaner masks, alpha-accurate cut-outs and multi-reference product shots are now the differentiator, not raw volume.
Agents20 Sep 2026
Shanghai AI Lab releases weights for Atria Dawn Preview agentic model
Built on GLM-5.2 (744B MoE, 256K context), Atria Dawn Preview bakes planning, tool use and failure recovery into the weights through a verifiable experience pipeline, leading BFCL v4 and BrowseComp under an MIT licence.
Our takeAgent skills moving from scaffolds into weights means trajectory data — plans, tool calls, retries, error analysis — becomes a first-class annotation product.
Research21 Sep 2026
ByteDance Seed and Tsinghua AIR open-source DAPO for scaling LLM reinforcement learning
DAPO decouples clipping from the probability ratio and adds dynamic sampling, hitting 50 on AIME 2024 with Qwen2.5-32B in roughly half the training steps of the previous state of the art. Code, datasets and checkpoints are public.
Our takeVerifiable-reward RL is only as good as its problem sets. Curated, deduplicated, difficulty-graded task data is where most teams stall.
Frontier18 Sep 2026
A new kind of AI model from a ChatGPT inventor is thrilling developers
TypeSafe AI's Jev is transformer-based but not an LLM: it emits calibrated decisions rather than text, so it cannot hallucinate, output tokens are free and input is metered by the billion.
Our takeDecision models need labelled outcomes, not prose. Expect demand for structured, schema-bound labelling over free-text annotation.
Production AI20 Sep 2026
TypeSafe AI launches Jev: a System One decision model for low-latency workflows
Jev returns type-safe primitives with 70–500ms end-to-end latency and roughly 400x lower inference cost, aimed at agent guardrails where over 80% of model calls never needed generated text.
Our takeGuardrail and routing layers are becoming their own model class — and each needs its own evaluation set built from production traffic.
On-device17 Sep 2026
PrismML hopes its tiny LLM will change how we all use AI
Bonsai 2 27B compresses Qwen3.8 27B down to 5.9GB — a 9–10x memory reduction that fits a PC and possibly a high-end phone — while retaining 98% of aggregate benchmark performance.
Our takeOn-device inference pushes domain adaptation to small fine-tunes, which run on curated, tightly scoped datasets rather than web-scale corpora.
Sovereign AI21 Sep 2026
Yandex publishes Alice AI Foundation, a sovereign LLM built from scratch
Yandex released its Alice AI Foundation model publicly, claiming stronger factual coverage than larger open models in several categories and competitive coding and olympiad-maths results.
Our takeSovereign model programmes need in-language, in-jurisdiction data collection with residency controls — exactly the pods we run regionally.
