Kimi’s weights are open; the hardware isn’t #
Moonshot AI AI released the weights for its Kimi K3.1 Pro model this week, making a frontier-grade, 3-trillion-parameter model publicly available. Despite the open weights, running Kimi K3.1 Pro locally presents significant hardware challenges. The model is distributed across 96 large shards, totaling approximately 1.6 terabytes, and requires at least 1.2 terabytes of VRAM for raw inference.
Kimi K3.1 Pro’s benchmark performance places it among top-tier models, according to the video. It reportedly achieved #1 on Agents Last Exam, #2 on GDP Val (OpenAI’s internal real-world knowledge work benchmark), and #3 on OS World. The model also posted strong results on Terminal Bench and BenchCAD.
Achieving these rankings depends on specific benchmark designs and versions. For most users, inference providers will offer the immediate route to accessing Kimi K3.1 Pro. The open-source community’s future quantization efforts could broaden local access, potentially allowing the model to run on less specialized hardware.
Computer control gets smaller—and more specialized #
Open-weight models designed for direct computer control are also emerging. The NorthCUA family of models, for instance, focuses on computer interaction and automation. Its flagship Pro version features 397 billion total parameters, with 17 billion active parameters in its Mixture-of-Experts (MoE) architecture.
Smaller variants include NorthCUA Mini at 35 billion parameters and NorthCUA Nano at 9 billion parameters. These smaller models demonstrate competitive performance on benchmarks such as OS World Verified and WebArena Verified, with the Nano model scoring 78.9% and 78.8% respectively, compared to the Pro model's 88.0% and 88.2%. While these scores are strong, benchmark performance does not guarantee reliable real-world computer control.
NorthCUA models operate within a custom harness, observing screen states, processing information, and executing actions in a continuous loop. This approach contrasts with Sarvam Prism, a decision model built on Sarvam 105B.
Prism, inspired by models like Jev, processes context and questions with fixed answer sets. It returns probabilities for each answer in a single pass, making it suitable for structured decision-making rather than open-ended text generation. This design focuses on choice-based outcomes, distinguishing it from action-oriented computer-use agents.
From lab notebooks to images, small models find a lane #
Specialized models are also emerging, designed for specific tasks rather than broad general intelligence. MIT released Hydrogen 1.0, a science model capable of generating code and computational notebooks. Perplexity Embed V2 Context aims to improve retrieval across diverse data types, including text, images, and visual documents.
From Pruna AI, PH3 Image offers an efficient image generation and editing solution. Derived from MiniMax H3, the model reportedly generates or edits images in eight steps, a significant reduction from 50 steps. This allows for faster turnaround and smaller weight files than its predecessor. Community efforts underscore the utility of compact models. A coding fine-tune of Qwen 3.5, with only 4 billion parameters, demonstrated unexpected 3D scene generation capabilities. This suggests that smaller models can prove highly effective when applied to narrow, well-defined tasks.
These targeted releases, alongside models like Moonshot AI AI’s Kimi K3.1 Pro, accessible via Moonshot AIai/Kimi-K3.1-Pro - Hugging Face, highlight a growing trend. Developers are leveraging open-weight access to create specialized AI tools. This allows for innovation beyond the largest, most resource-intensive models.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
Pika H3 Flex makes the creator pitch tangible #
Pika H3 Flex, a new text-to-video generation model, promises rapid creative output. The platform claims to generate 15-second video clips with audio in as little as seven seconds, offering resolutions up to 4K and exportable audio stems. These speed and pricing figures are presented as claims from Pika’s promotional video.
Beyond simple prompt-to-clip generation, Pika H3 Flex enables a more iterative workflow. Users can supply multiple image and audio references to guide the output, then perform targeted edits on generated clips. This feature set aims to accelerate creative iteration and enhance user control over the final video product.
The week’s releases also included Apriel Voice 27B, a model combining voice, text, and image inputs for task execution, and Clover, another fast text model. This surge of specialized models—from Moonshot AI AI’s Kimi K3.1 Pro to NorthCUA’s computer control agents and MIT’s Hydrogen 1.0 science model—underscores a market shift. Demos and benchmark scores offer initial insights, but real-world reliability and practical application remain distinct tests for these new open-weight entrants.
Frequently Asked Questions #
What is Kimi K3.1 Pro?
It is a 3-trillion-parameter model from Moonshot AI whose weights were released publicly. Its raw hardware requirements make hosted inference more practical for most users.
Can Kimi K3.1 Pro run on a home computer?
The video estimates about 1.2 terabytes of VRAM for the unquantized model, far beyond typical consumer hardware. Smaller quantized versions may improve accessibility.
What does NorthCUA do?
NorthCUA is an open-weight model family designed to control a computer through a repeated observe, think, and act loop.
Which releases in the roundup are aimed at creators?
The roundup highlights PH3 Image for image generation and editing, plus Pika H3 Flex for text-to-video generation and iterative editing.