Kimi K3.1 weights OUT, open source computer use, SOTA image model, science model, Pika H3: AI NEWS
Matthew Berman showed 13 AI tools in this video on . 12 are verified available in the EU. Each row below jumps to the exact second. Part of week 2026-45.
Tools shown in this video, with timestamps
Moonshot released open weights for Kimi K3.1 Pro, a roughly 3T-parameter frontier-grade model needing about 1.2 TB VRAM. It ranks near the top on Terminal Bench, Agent SL Exam, OS World and GDP Eval, just behind Opus 5.5 in several.
Open-weights computer-use model family: Pro (397B total/17B active MoE), Mini (35B) and Nano (9B), with an open-source harness. It scores 78.9/85/88 on OS World Verified for Nano/Mini/Pro.
Pruna AI turned the MiniMax H3 video model into an open-source image generation and editing model. It runs in 8 steps instead of 50, is about 7x smaller, and makes a 4MP image in 2.6s. It is strong at photorealism and handwriting.
Live demoOpen-weights decision model built on Sarvam 105B. Given context and questions with fixed answer options, it returns a probability for each answer in one forward pass without generating text. It needs about 106 GB of disk and a B200/H200-class GPU.
A scientific-discovery model post-trained from GLM 5.3. It scores 38.33% on Frontier Science at $0.25 per problem versus $0.88 for Astra, and the demo wrote code to simulate a polyethylene melt in 3D. Weights are on Hugging Face and an API is available.
Live demoOpen-weights vision-language embedding models in 9B and 600M sizes, for retrieval over text, images and visual documents such as PDFs. A paper was released alongside it.
A community fine-tune of Qwen 3.5 4B that generates standalone three.js HTML/JavaScript 3D scenes with GLB export, useful for editors like Godot. Examples shown include jungle and snow siege platforms.
Live demoOpen-weight text-to-video model trained on MiniMax H3, ranked number one open-weight on Design Arena. It supports up to 4K, audio stems and multi-asset inputs, makes a 15-second clip in about 7 seconds, and costs about 1 cent per second. The creator says he uses it locally.
Live demoOpen-weights 27B voice-first multimodal MoE model that takes speech, text and images and acts on them, including text-to-speech output. It scores well for its size on benchmarks.
Open-weights diffusion language model with 35B total and 3B active parameters. It scores well against similar-size models. In a computer-use demo it finished a task in 169 seconds while Qwen 3.6 Flash hit its step limit at 578 seconds without completing.
Live demoNew generation of Liquid foundation models for on-device use. The 1.7B version outperforms the previous 2.6B, fits in about 1 GB at 4-bit, has a 256K context, and handles video, text and reasoning.
Open-source OCR model with no page limits or growing attention cost. It parses very long documents in one pass and held up on 5,000-page tests. The creator says he tested it and found it impressive.
Closed OpenAI model with 50% lower standard API prices than GPT 5.6 Terra and better capability. It matches GPT 6 Astra at about 1/10 the per-token price and finishes tasks over 3x faster, per the cited charts.
