ModelsNews
On October 19 Inception released Mercury 3, its most capable diffusion language model so far. The company reports 1,068 tokens per second on NVIDIA GPUs and a 1 million token context. The OpenAI-compatible API is also on OpenRouter, Baseten and Vercel.
· 2 min read
ModelsNews
China's MiniMax released M3.1 Flash on Monday. The company says it is a generational leap in agentic tasks, coding, computer use and cybersecurity. Online inference is nearly three times faster, and usage costs are 70% lower than its predecessor.
· 3 min read
ModelsNews
doubleAI, the new company of former Mobileye CEO Amnon Shashua, has unveiled its first product. Maxwell is what doubleAI calls a code execution model: instead of writing code, it speeds up running programs and formally proves every change equivalent to the original.
· 2 min read
ModelsAnalysis
We asked two questions. Which AI model gives you the most intelligence per dollar? Claude Haiku 5.5 wins with 217 index points per dollar. Which of the five smartest models is cheapest? Claude Sonnet 5.5, at a blended price of $4.
· 5 min read
ModelsNews
Paper Instruments unveiled Ultramarine 1 Flash, an agentic model tuned for professional knowledge work. It builds on the open GLM-5.3 Flash model and, on the company's internal benchmark, costs $0.13 per task while outscoring Claude Fable 5.
· 2 min read
ModelsNews
Mistral released Timestral, its first family of time-series foundation models. The largest variant has 209M parameters and, according to Mistral, ranks first on all three main public forecasting benchmarks. Inference does not need a GPU.
· 2 min read
ModelsNews
China's StepFun released open weights for Step 5 Preview on October 15, its flagship model for agentic work. It is a sparse MoE with 600B total parameters, 27B active per token and a 1M-token context window.
· 2 min read
ModelsNews
Perplexity's Decisions API now serves pplx-decider-v1.2 and a new 4B Flash model. Both return probabilities for text and images instead of generating text, at $0.02 and $0.01 per million input tokens.
· 4 min read
ModelsNews
Liquid AI released d1-ultra, a multimodal decision model. It returns probabilities from text and images instead of generating text. It costs $0.042 per million input tokens and is live on the Liquid API.
· 4 min read
ModelsNews
DeepSeek released an updated version of its V4.1-Flash model on October 17. According to the company, DeepSeek-V4.1-Flash-1017 improves on all 11 reported benchmarks, finishes tasks faster and produces shorter outputs. It is live on the web, app and API, and the weights are available under the MIT License.
· 4 min read
ModelsNews
Elorian unveiled Glimpse on October 16, calling it a “visual thinking model” that reasons about the physical world from images and video. The company says it matches GPT-6 Astra on several multimodal benchmarks while costing 94% less per token.
· 2 min read
ModelsNews
Black Forest Labs (BFL) released FLUX 3 Image Klein, an open-weight image model, on October 15. The 5.6-billion-parameter model handles text-to-image, image editing and generation with up to 10 reference images. Its INT8 build runs on GeForce RTX cards with 8 GB of VRAM.
· 3 min read
ModelsNews
US lab Reflection released its first open-weight model on Hugging Face on October 16. Beam is a 501-billion-parameter mixture-of-experts (MoE) model that uses 23 billion parameters per token. It targets coding and agentic work, handles a 1-million-token context and ships under the Apache 2.0 license.
· 4 min read
ModelsNews
NVIDIA released Nemotron 3.5 Super VL on October 15: an open 120-billion-parameter multimodal model that activates 12 billion parameters per token. Two smaller Bolt models arrive alongside it to run agents on local hardware.
· 2 min read
ModelsNews
Microsoft AI said on October 15 that it has finished pre-training MAI-Base-2. The 2.8-trillion-parameter MoE base model activates 80 billion parameters per token and, according to Microsoft's own held-out tests, beats every open-weight model it evaluated.
· 2 min read
ModelsNews
Abu Dhabi's Institute of Foundation Models (MBZUAI) and Inception42 released Jais 3 on October 15: six open Arabic language models from 900 million to 375 billion parameters, plus the Jais Speech recognition model.
· 2 min read
ModelsAnalysis
A quiet week for the big Chinese labs, so we take stock: GLM-5.3-Flash pricing, the gap to US models on independent evals, and how to use these models GDPR-compliantly from the EU.
· 8 min read
ModelsAnalysis
Google made Gemini 4 Argon, the first Gemini 4 model, generally available on October 13. It brings a 1M-token unified input-output window, introductory pricing of $2/$10 per million tokens, and strong agentic and long-context results. Independent tests put Google back among the leaders, but the picture is more nuanced than the vendor's table.
· 9 min read
ModelsNews
Meta released the weights for Muse Spark Open on October 15: a natively multimodal model with 816 billion total parameters (42 billion active) and a 1-million-token context window. Meta calls it the most capable US open-weight model to date, a claim based on its own evaluations.
· 3 min read
ModelsAnalysis
TypeSafe's Jev doesn't write text. It returns typed decisions with probabilities, at $0.042 per million input tokens, with output free. Its "can't hallucinate" claim is really a schema guarantee: the answer is always one of the options you define, but it can still be wrong.
· 5 min read
