Kimi K3.1 weights OUT, open source computer use, SOTA image model, science model, Pika H3: AI NEWS

Summary

The video surveys a large batch of AI model releases, emphasizing open weights, specialized capabilities, efficiency, and models that can run locally. Benchmark results and quality comparisons below are claims presented in the video.

Model releases

• Moonshot Kimi K3.1 Pro: Moonshot released the weights for a large frontier model with roughly 3 trillion parameters. The presenter says it performs strongly on agent, coding, computer-use, and knowledge-work benchmarks. The weights are available on Hugging Face, but the full model is too large for most local setups; the video estimates that it would require around 1.2 TB of VRAM. The presenter expects community-made quantized versions to make it more accessible.

• NorthCUA: An open-weight computer-use model family with Pro (397B total / 17B active parameters), Mini (35B), and Nano (9B) versions, plus an open-source harness. The video highlights strong OSWorld Verified scores and a relatively small performance gap between Nano and Pro. The agent repeatedly observes screenshots, reasons, and takes actions.

• Pruna P-H3-Image: An image-generation and editing model derived from MiniMax H3. The release is presented as a faster, smaller alternative: it reportedly generates or edits images in 8 steps rather than 50, with a 4-megapixel image taking about 2.6 seconds in the cited example. The presenter shows results for photorealistic, anime-style, and detailed handwritten-looking images. Weights are available on Hugging Face.

• Sarvam PRISM: An open-weight decision model, rather than a text-generation model. Given context and questions with fixed answer choices, it returns probabilities for each option in one forward pass. The video notes that its recommended hardware, such as B200 or H200 GPUs, is not typical consumer equipment.

• Hydrogen 1.0: A scientific-discovery model created by post-training GLM 5.3. The presenter shows it writing code to simulate a polyethylene melt at the molecular level, with a 3D visualization. The video reports a Frontier Science score of 38.33% and a cost of $0.25 per problem, and says its benchmark results exceed those of the compared models shown. Weights and API access are mentioned.

• Perplexity Embed v2 Context: An open-weight vision-language embedding model for retrieval across text, images, and visual documents such as PDFs. It comes in 9B and 600M parameter versions; the smaller option is described as suitable for local use.

• Qwen3.5 4B 3D Coder: A community fine-tune of a 4-billion-parameter model focused on 3D coding. It generates standalone Three.js HTML/JavaScript scenes and can export GLB assets. The presenter suggests it could be useful for indie game developers and asset creation.

• Pika H3 Flex: A text-to-video model based on MiniMax H3. The presenter describes it as a leading open-weight option for video generation, highlighting camera control, audio stems, resolutions up to 4K, and fast generation. The video says a 15-second clip with a soundtrack can be produced in as little as 7 seconds, with pricing starting at about 1 cent per second. It can also use multiple image and audio inputs and supports further edits.

• Apriel Voice 27B: An open-weight, voice-focused model from ServiceNow. It can process speech, text, and images together, interpret what a user wants done, and take action. It also supports text-to-speech. The video describes it as a relatively compact model that can run locally on suitable hardware.

• DiffuSpace Clover: A 35B-parameter mixture-of-experts diffusion text model with around 3B active parameters. The presenter emphasizes its speed, comparing a computer-use task that Clover completed in 169 seconds with a longer-running model that hit its step limit without finishing.

• Liquid AI LFM3: A small, on-device-oriented model family. The video says the 1.7B version outperforms an earlier 2.6B model and can fit in roughly 1 GB at 4-bit precision. It is described as supporting a 256K-token context window, local agents, video, text, spatial understanding, and step-by-step reasoning. Fine-tuning on a single GPU is also mentioned.

• Baidu Unlimited OCR 2: An open-source OCR model for extracting and understanding text in image-based documents. The video says it can process very long documents in one pass, citing a test on 5,000-page documents. The presenter reports trying it and finding it impressive.

• GPT-6.1 Terra: The only non-open-weight model covered. The presenter calls it a strong-value option, citing lower API prices than its predecessor, faster task completion, and strong agent-task results relative to more expensive models. Team members are also said to have found it capable at video editing.

Guides and practical notes

The video is primarily a news roundup with brief demonstrations and setup pointers; it does not present a step-by-step tutorial.

• NorthCUA: Its Hugging Face page is said to include setup instructions and a link to the open-source harness. The presenter suggests asking an AI agent to install it using those instructions.
• Model access: For Kimi K3.1 Pro, P-H3-Image, PRISM, Hydrogen, Qwen3.5 4B 3D Coder, Pika H3 Flex, Apriel Voice, Clover, LFM3, and Unlimited OCR 2, the video points viewers to model pages, papers, or project links for downloads and further details. Some large models will require inference providers or substantial GPU hardware rather than ordinary local machines.

Main speaker and sources

• Main speaker: Matthew Berman
• Sources discussed: Moonshot AI, Beipo AI (NorthCUA), Pruna AI, Sarvam AI, Deep Principle (Hydrogen), Perplexity, Qwen contributors, Pika, ServiceNow AI, DiffuSpace AI, Liquid AI, Baidu Research, and OpenAI.