Model launchesGoogle unveils Gemini 4 Carbon model focused on cybersecurity and software tasksGoogle announced Gemini 4 Carbon, a top-tier AI model optimized for software development and cybersecurity workloads.2 min read
Model launchesMistral publishes weights for trillion-parameter Mistral Large 4, highlights cybersecurity usesFrench AI developer Mistral has released the model weights for its trillion-parameter Mistral Large 4 (ML4), nicknamed “Le Chonk,” enabling organisations to download and run the model on their own infrastructure.3 min read
Model launchesMistral Large 4 (Le Chonk): an open-weight multimodal MoE model with 1M-token contextMistral Large 4 (aka Le Chonk) is a multimodal, multilingual Mixture‑of‑Experts model released by Mistral with about 1T total and 49B active parameters and a 1,000,000‑token context window.5 min read
Model launchesNvidia releases open-weight Nemotron 3.5 Transcribe and Granary v2 speech datasetNvidia published Nemotron 3.5 Transcribe, an open-weight speech transcription model that handles recorded audio in up to 40 languages and claims support for 31 languages to and from English, hour-long recordings, and up to eight speakers.2 min read
Model launchesHow Anthropic ships Claude: separating the brain from the handsAnthropic’s recent S‑1 disclosures and a company podcast outline how the lab balances model quality, cost and safety in shipping Claude—centered on the Opus 5.5 launch, Managed Agents architecture, and platform engineering choices.7 min read
Model launchesOpenAI launches mid-tier GPT-6.1 Terra focused on visual reasoning and computer useOpenAI introduced GPT-6.1 Terra, a mid-range model emphasizing detailed visual reasoning and precise computer interaction while producing clearer, more natural text.2 min read
Model launchesDeepSeek V4.1 Flash available on the Cursor platformThe DeepSeek V4.1 Flash model is now available on the Cursor interface; the article says every effort level falls within the CursorBench cost threshold, task costs are listed as $0.60–$0.98, and other models do not perform better.1 min read
Model launchesSingle Researcher at Mirendil Trains Large Coding Model, Demonstrating Faster AI R&DAI R&D startup Mirendil announced Mirendil‑Base, a pretrained coding model developed by a single researcher in about 3.5 months using under $10 million of compute.2 min read
Model launchesBlock unveils Dessa, a human-like full‑duplex speech model that runs locallyBlock, Jack Dorsey’s company, introduced Dessa on October 23, 2026 — a full‑duplex speech‑to‑speech model trained entirely on human speech that runs on a single NVIDIA GPU and can operate locally on consumer machines.2 min read
Model launchesMoonshot unveils Kimi K3.1 Pro — 2.8T MoE model with promised open weights and faster APIMoonshot announced Kimi K3.1 Pro, a multimodal mixture‑of‑experts model with 2.8 trillion total parameters (104B active), a 1‑million‑token context window, and a promise to publish weights on November 4, 2026.5 min read
Model launchesMuse Wonder Full judged effective for orchestration and skill-writingAn evaluation of the model called Muse Wonder Full states that the user regularly employs it for orchestration, rewriting documents and skill descriptions, and steering models.1 min read
Model launchesCanva launches Canva Craft image model focused on design-aware, layered outputsCanva introduced Canva Craft, its first in-house image-generation model built from Canva’s library of human-designed templates.2 min read