AI News: Kimi K3.1 Weights, Computer Use and Specialized Open Models

This week brought a wide range of model releases: open weights for Kimi K3.1 Pro, a computer-use model, image and scientific models, and several smaller specialized systems. I will walk through the examples and reported benchmarks, which should be read in the context of each developer’s test setup.

Moonshot has released [Kimi K3.1 Pro weights](https://huggingface.co/moonshotai/Kimi-K3.1-Pro). I consider this a strong open-weight model, but it is very large: the files shown are split across 96 parts of roughly 17 GB each, so an ordinary local machine is unlikely to run the full version.

I estimate that running the model could require more than a terabyte of VRAM, depending on precision and deployment, and expect the community to explore quantized variants. The exact hardware requirement depends on the implementation; open weights do not automatically mean an unrestricted open-source license.

The published chart compares Kimi with several closed models on terminal, agent and other benchmarks. These results are useful comparisons for the tested configurations, but a high score on a particular benchmark does not establish the best model for every task.

[NorthCUA](https://huggingface.co/beipo-ai/NorthCUA-Pro) is a computer-use model family from Beipo AI. I describe the Pro version as a mixture-of-experts model with 397 billion total and 17 billion active parameters, alongside smaller Mini and Nano versions.

The team also provides a harness. In the desktop-agent loop I show, the system observes a screenshot, decides what to do and takes an action, then repeats. I quote OSWorld Verified results from the release; those are reported scores, not my independent reproduction.

Pruna AI’s P-H3-Image adapts MiniMax H3 for image generation and editing. The demonstration compares eight generation steps against 50 and claims faster output with a smaller model; measured speed and quality depend on hardware and the chosen comparison.

The examples include photorealistic and anime-style images and a handwritten-looking recipe. I prefer some of the displayed outputs, but those are subjective judgments based on selected examples rather than a broad quality evaluation.

[Sarvam PRISM](https://huggingface.co/sarvamai/prism) is presented as a decision model rather than a text generator. Given context and fixed answer options, it returns probabilities for those options; that makes its interface different from asking a chat model to write an answer.

I note that the model weights occupy substantial disk space and the model card recommends data-center-class GPUs. Whether a particular local machine can run it depends on memory, precision and the available implementation.

Hydrogen 1.0 is presented as a model for scientific work, post-trained from GLM 5.3. I show a Frontier Science comparison and a reported cost per problem. Those numbers come from the release’s chosen benchmark and do not by themselves establish scientific reliability.

The demonstration asks for a polyethylene-melt simulation and shows code and a molecular visualization. A model-generated simulation is an example of a workflow; its assumptions and results still need domain-expert validation.

Perplexity Embed v2 Context is described as an open-weight vision-language embedding family in 9B and 600M sizes, aimed at retrieval across text, images and visual documents. I point viewers to the accompanying paper and model details.

A community fine-tune of Qwen3.5 4B targets 3D coding, including Three.js scenes and GLB export. I show a small game-style example and suggest such a specialized small model could be useful for asset prototyping; quality will vary by task.

Pika H3 Flex is a video model based on MiniMax H3. I show sample clips and the developer’s claims about output speed, resolution and pricing; the price and performance depend on service terms and configuration.

In the demonstrated editing flow, several images and an audio file are combined into a generated clip, with later changes to individual elements. I find the examples promising, but they are selected demonstrations rather than a controlled comparison.

ServiceNow’s [Apriel Voice 27B preview](https://huggingface.co/ServiceNow-AI/Apriel-Voice-27B-preview) is described as accepting speech, text and images for task-oriented interactions, with speech output. I discuss its reported benchmarks; running it locally still depends on available hardware and precision.

DiffuSpace Clover is a diffusion-based model described as 35B total and 3B active parameters. I show a computer-use example in which it completes the demonstrated task faster than a comparison model; one example does not measure performance on every task.

Liquid AI’s [LFM3](https://www.liquid.ai/blog/lfm3) targets on-device use. I highlight its small model size, a 256K context claim and reported improvements over an earlier version. Memory usage and attainable speed will vary by device and configuration.

Baidu’s [Unlimited OCR 2](https://github.com/baidu/Unlimited-OCR) is a document OCR system. I discuss its reported long-document tests and my positive impression; accuracy can still vary with document quality and layout.

Finally I discuss OpenAI’s GPT-6.1 Terra, the closed-weight model in this roundup. I relay the stated price and benchmark comparisons and my colleagues’ positive impressions of video editing. These are time-specific claims, and actual costs depend on the workload and current pricing.