Open-Weights Week: Kimi K3.1 Pro, NorthKa, Pruna's PH3, Sarvam Prism, Hydrogen 1.0, Pika H3 Flex and More

Berman opens with Moonshot's release of the Kimi K3.1 Pro weights, which he calls an actual frontier model. He says it is not at the level of the absolute frontier from closed labs such as Anthropic and OpenAI, but is basically a half step behind them, and that it is completely open source and open weights. He notes that Moonshot had promised the release and that the weights can now be downloaded from Hugging Face.

He cautions that the model is not small and that most people are unlikely to run it locally, so they would need to use an inference provider. The Hugging Face listing shows 96 parts of roughly 17 GB each. Berman estimates that running the model would require at least 1.2 TB of VRAM, which he calls a ton. He says the community will likely produce quantized versions that could eventually run on local machines.

Kimi K3.1 Pro is a 3 trillion parameter model. Berman says he believes the absolute frontier models such as Astra and Fable are closing in on 10 trillion parameters, which makes Kimi a little less than half their size.

Berman walks through the benchmark results. On Terminal Bench, Kimi K3.1 Pro beats Astra and sits right below Opus 5.5, which he calls the current best model on the planet. On Agents Last Exam it scores number one, ahead of Opus, Astra and the others. On Bench CAD, which tests the ability to create CAD files, he says it is quite good but falls behind Opus 5.5 and Astra.

On OSWorld, which measures computer use, it ranks third behind Fable and Opus. On GDPval, which he describes as OpenAI's internal model for real-world knowledge work, it scores second on the ELO score, right behind Opus 5.5. Berman calls it a phenomenal model and says he is very excited about the quantized versions that are expected soon.

Next Berman covers NorthKa, a medium-sized computer-use model. The flagship has 397 billion total parameters with 17 billion active, making it a mixture-of-experts model that he says is extremely good at computer control. The family also includes NorthKa Mini at 35 billion total parameters and NorthKa Nano at 9 billion total parameters. An open-source custom harness for running the models, built by Beo AI, is also available.

On the OSWorld Verified computer-control benchmark, Nano scores 78.9, Mini scores 85 and Pro scores 88. Berman points out that the gap between Nano and Pro is only about 10 points, and that the 9 billion parameter model can run on most hardware, perhaps even a phone. He also notes close scores between Nano and Pro on another verified benchmark, which he describes as strong frontier-grade results.

In the harness, the desktop agent works by taking screenshots, observing what it sees, thinking, and then acting, repeating that loop over and over. Berman says the Hugging Face page includes setup instructions and a link to the harness, and suggests that users can simply give the link to their agent and ask it to set the system up.

Berman notes that many projects are being built on MiniMax H3, a video model, but that Pruna AI took it and made it into an image model. Pruna has open-sourced PH3, an image generation and editing model built from MiniMax H3. Berman says the company took the video model and turned it into a state-of-the-art image model, with a focus on efficiency and speed.

On efficiency, Berman says the model generates or edits images about five times faster, using eight steps instead of 50. A 4-megapixel image takes 2.6 seconds instead of 12.4 seconds. He also says the model is seven times smaller than H3, so high-resolution images can be downloaded easily and run on a much smaller GPU.

Berman lists its strengths as photorealism, documentary and editorial photography, film stills, and graphic styles with short quoted text. The company released its full technical report on its blog.

Berman shows examples, starting with a prompt for an old fisherman at sunrise. He says the image looks pretty good and becomes more photorealistic with the image upsampler from PH3. A robot gardener in anime style also looks good, and he says a second version looks much better and more in the style of anime, as he understands it.

He gives file sizes: MiniMax H3 is 143 GB, while PH3 image is 55 GB at BF16 and 30 GB at int8. In a comparison at 50 steps for H3 and eight steps for PH3, he says the images with aliens and humans in space are almost identical. He says the other comparison images look nearly identical while taking far fewer steps and being much more efficient. He also praises the model for handwritten-style images, showing a recipe with a coffee-stain effect and examples of pencil and ink writing with some mistakes, and says the weights are open on Hugging Face.

Berman introduces Sarvam Prism as an open-weights, open-source counterpart to a decision model he had praised earlier, which he compares to Jev. Prism is built on Sarvam 105B. A user provides a context and one or more questions, each with a fixed set of possible answers, and the model returns a probability for every answer in a single forward pass. Berman stresses that it does not generate text, so the answer is always one of the given options.

The model recommends a B200 or H200 GPU, which Berman notes are not necessarily consumer-grade chips, though he thinks it might run on a fairly beefy consumer machine. It requires 106 GB of disk space for the weights, and the Hugging Face page provides the setup steps.

Berman next covers Hydrogen 1.0 from MIT, a model built specifically for scientific discovery. MIT post-trained GLM 5.3, an open-source open-weights model from a Chinese lab, and made it stronger at scientific discovery. On the Frontier Science benchmark, he says Hydrogen scores 38.33%, and beats GLM 5.3, Kimi K3, and GPT-6 Astra Max, while being less expensive than almost all of them, with the exception of GLM 5.3. He gives a cost of 25 cents per problem versus 88 cents per problem for Astra.

As an example, Berman shows the model simulating a polyethylene melt. He describes it as zooming into molecules, rendering them in 3D space that can be rotated, and writing the code for the simulation at the molecular level. The output shows metrics including density, RDF, radius of gyration and end-to-end distance. He notes that the weights can be downloaded from Hugging Face or used through an API.

Berman says Perplexity has dropped an open-weights embedding model and thanks the company for contributing to the open-source community. The release, Perplexity Embed V2, is an open-weights vision-language embedding model available in a 9 billion parameter version and a 600 million parameter version. A paper accompanies it, which Berman does not go over but links in the description.

He describes it as a state-of-the-art vision-language embedding model for retrieval over text, images and visual documents such as PDFs, meaning it is good at finding text, images and content within a document. He says it is open source and open weights and can be run locally on a computer.

Berman calls Qwen 3.5 4B a really small, really capable coding model with 4 billion parameters that is especially good at 3D coding. It generates standalone Three.js HTML and JavaScript scenes with GLB export, which he says makes it useful for editors such as Godot. He shows a 3D scene it created and says he likes having small, specialized models that can run quickly on a local machine.

He shows additional outputs, including a jungle siege platform and a weathered snow siege platform. Berman suggests that indie game developers could use the model to create some of their 3D assets.

Berman calls Pika H3 Flex possibly the best text-to-video model he has ever used. He notes that it is trained on MiniMax H3, but he says it is faster and better overall, and that he has been using it locally quite a bit. Pika debuted it as the number one open-weights model on Design Arena's text-to-video board. It is built for professional iteration, with exportable audio stems, resolutions up to 4K, and generation faster than real time.

According to Berman, it can produce a 15-second clip with a soundtrack in as little as seven seconds, and he says it is inexpensive, starting at one penny per second, so a 60-second video costs about 60 cents. He shows examples including two players playing ping-pong in a pool, one of whom is a frog, and an intro with lots of camera movement and realistic hair. He also demonstrates giving the model multiple assets, five images and a thunderstorm audio file, which it combines into one video. Outputs can be edited afterward, for instance by changing one part of the video.

Berman says he did not expect ServiceNow to appear here, but introduces an open-source, open-weights model that he says is good at understanding voice and turning it into real-world tasks. The model is Apriel Voice 27B. He describes it as a relatively small model that can run on most local machines, and as a voice AI model built to get work done.

Berman says the model can take in text, images and voice together, combining them and acting on what it understands the user wants to get done. He lists its components as a speech encoder, token embeddings and a frozen visual interface, all fed into a mixture-of-experts backbone that decides what to do. The model can also perform text-to-speech. Berman says that for its size it scores very well on the benchmarks and that it is very voice first.

Berman introduces Clover from Diffuse Space AI as another diffusion text model, which he says should bring very fast speeds. It is a 35 billion parameter mixture-of-experts model with 3 billion active parameters that scores strongly against models of similar size. To show its speed, he runs a prompt asking how long it takes to drive from CMU Pittsburgh to the nearest Wendy's.

He says Clover finished the task after 169 seconds. By contrast, Qwen 3.6 Flash kept going and reached its 30-step limit at 578 seconds without completing the task, after 28 steps in total. Berman says he was impressed by how fast Clover operated the computer.

Berman presents LFM3 from Liquid AI, a new generation of liquid foundation models built for on-device deployment, which he says are not autoregressive language models. The LFM3 1.7B model outperforms the previous 2.6 billion parameter version while using 35% fewer parameters. At 4-bit precision it fits in about 1 GB, so he says it could run on any device including a phone. He adds that it is at the Pareto frontier compared with models of its class and size.

Berman covers Unlimited OCR 2, an OCR model that reads image-based documents, which he explains means reading text from images and understanding them. The release is open source, with no page limits and no growing attention costs, and it parses documents of virtually unlimited length in one go while maintaining high accuracy and efficiency. Berman says it was tested on 5,000-page documents and held up well, performs very well on most OCR benchmarks, and that he tested it himself and found it impressive.

Berman closes with GPT-6.1 Terra from OpenAI, which he describes as the only non-open-source, non-open-weights model in the video, while noting that it is worth discussing because it is apparently very capable. He says two members of his team told him it is very good at video editing. Compared with GPT-5.6 Terra, the previous version, Terra has standard API prices that are 50% lower and is also more capable.

On the Agents Last Exam cost-performance chart, Berman says Terra sits at almost zero cost, and that GPT-6 Luna is the only model less expensive overall. At the lowest thinking budget, he says Terra costs basically the same as Luna but delivers a substantial improvement in intelligence. He notes that Astra at its highest thinking budget scored similarly to Terra at its highest, but Terra costs a fraction of the price, which he calls very economical for massive workflows that use many tokens.

Berman says Terra matches GPT-6 Astra at one-tenth the price per token and finishes tasks over three times faster on average than both Astra and GPT-6.1 Sold. He calls it super impressive and encourages viewers to try it.