Matthew Berman's AI news roundup covers thirteen releases in under twenty minutes. The week is notable for available model weights, but “open weights” does not mean “runs on my laptop,” “licensed for my country,” or “independently proven best.” The useful way to read these releases is by job: computer use, scientific assistance, media creation, retrieval, voice, OCR and inexpensive everyday inference.
Start small and task-first. LFM3-1.7B is the clearest on-device text experiment; Perplexity's 0.6B embedding model is for document retrieval, not chat; CCORE's Qwen-derived 3D coder is a focused coding test. The full Kimi K3.1 Pro repository is about 1.56 TB. For readers in Portugal or elsewhere in the EU, the published P-H3-Image and Pika H3 Flex weights carry a MiniMax H3 license that excludes EU use without a separate license.
Watch Matthew's Roundup
Source note, 6 November 2026: Matthew's reactions and demos come from the video and supplied transcript. We checked the linked publisher model cards, repositories, technical posts and pricing pages for availability, scope and limits. We did not run the models or reproduce vendor benchmarks. A posted checkpoint proves access to weights, not training-data openness, easy local deployment, performance or permission to use it everywhere.
The 13 Releases, Organized by Job
Each link goes to a publisher's model card, repository or release page. “Try” means a route to investigate after checking current terms; it is not a blanket install recommendation.
| Release and source | What it does | Practical check |
|---|---|---|
| Kimi K3.1 Pro | Large multimodal coding and agent model. | 2.8T total parameters and about 1.56 TB of files. Use hosted access for a first test; read its custom license. |
| NorthCUA family | Screenshot-driven desktop and browser control. | Pro, Mini and Nano have different sizes; the official harness is part of the workflow. Test only in a disposable, permission-limited environment. |
| Pruna P-H3-Image | Image generation and editing from MiniMax H3. | Fast vendor results use a B200; quantized ComfyUI files need about 21 GB VRAM. Its model license excludes EU use without separate permission. |
| Sarvam PRISM | Score fixed choices in a decision task. | Not a chat model. Publisher asks for about 106 GB disk and an FP8-capable GPU with roughly 120 GB free memory; a hosted API is available. |
| Hydrogen-1.0 | Assist scientific reasoning and research workflows. | Built by Deep Principle on GLM-5.3, not by MIT. The publisher's examples use external tools; scientific conclusions still need expert verification. |
| pplx-embed-v2-context | Search text, images and visual documents with contextual embeddings. | 0.6B and 9B weights exist. They produce vectors for retrieval, not answers; test on your own documents and queries. |
| Qwen3.5 4B 3D Coder | Generate editable Three.js scenes and 3D code. | A CCORE community fine-tune, not an official Qwen release. Validate scene code, export and frame rate rather than judging screenshots alone. |
| Pika H3 Flex | Iterative video and synchronized audio generation. | Local weights cover 480p/768p; 4K finishing and stems are hosted features. The MiniMax H3 weight license excludes EU use without separate permission. |
| Apriel Voice 27B Preview | Understand speech, reason and emit tool calls. | The checkpoint outputs text and tool calls. Spoken replies require a separate TTS system; do not assume the model alone is a full-duplex voice assistant. |
| Clover 35B-A3B Preview | Diffusion-language coding and tool use. | Speed and benchmark numbers are publisher tests on specific GPUs. The release notes call it a preview; measure actual task latency and quality yourself. |
| Liquid AI LFM3 | Small specialist models for on-device text, vision, audio and decisions. | It is a family; the 1.7B text model is not every modality at once. Liquid says its 4-bit text version fits in about 1 GB. |
| Unlimited OCR 2 | Parse long image-based documents with cross-page context. | The model card gives multi-page code tested on an A800 80 GB. “Unlimited” and 5,000-page examples are research claims, not a guarantee for every scan. |
| GPT-6.1 Terra | Closed-model visual reasoning and computer use. | Available through OpenAI products and API, not downloadable weights. Standard API rates are $1 input / $5 output per million tokens; check current rates before a cost comparison. |
Agents, Decisions and Science Need Different Tests
Kimi: The 96-shard repository lists about 1.56 TB of files. That is a download footprint, not a universal VRAM requirement: serving memory also depends on quantization, runtime, context, offloading and throughput. For most readers, the realistic first experiment is through Kimi or a compatible hosted provider, using a representative coding task and a cost log. Moonshot calls the model open-weight but uses its own Kimi K3.1 Pro License, which has special terms for large model-as-a-service businesses and very large products. Avoid equating access to weights with a standard permissive open-source license.
NorthCUA: Beipo AI's Pro card describes the model and harness as a pair. The screenshot-to-action loop is the product, so a model-only benchmark will miss the tool, coordinate and recovery behavior. The smaller Nano is a more approachable research target than Pro; that does not establish that it will run well on a phone. Keep the first run on a throwaway desktop with no real credentials or payment access.
PRISM and Hydrogen: PRISM emits probabilities over specified options in one pass. The question is calibration on your decision set, not prose quality. Hydrogen-1.0 comes from Deep Principle, which post-trained GLM-5.3 for scientific work. Its model card says demonstrations involve external tools and that outputs require review against the scientific evidence. A generated notebook or simulation is a hypothesis until its assumptions and results are checked.
Image, Video and 3D: The License Comes First
Pruna's P-H3-Image is a real image/editing release, and Pruna reports a large speed improvement over its H3 baseline. The headline 2.6-second result is measured on a B200, while the model card gives very different memory figures for full-precision and quantized use. It also lists limitations around hands, background faces and dense text. Most importantly for this site's European readers, its MiniMax H3 Community License excludes the EU, UK, Republic of Korea and US unless a separate license is obtained. A public download button does not override that term.
Pika H3 Flex also inherits the MiniMax H3 territory restriction. Its launch post is explicit: local code and adapters cover native 480p/768p video, while 1080p-to-4K finishing and separate audio stems are Pika-hosted features. The “seven seconds for fifteen seconds” statement measures transformer computation, not total API turnaround. Matthew also says one cent per output second and then calls a sixty-second video $60; that arithmetic would be $0.60 at that rate, before resolution, references, finishing and clip limits. Use Pika's current task-specific pricing for an actual budget.
CCORE's Qwen3.5 4B 3D Coder is the lightweight creative experiment here. Its community model card shows standalone 3D scenes; those are not proof of clean assets, accessible interactions or a game-ready project. Ask it to generate one small scene, run it in a browser, inspect geometry and code, and compare against the base model on the same prompt.
Retrieval, Voice, OCR and Small Local Models
Perplexity Embed v2 Context is a retrieval component. Its 0.6B and 9B variants encode text and visual documents into vectors so a search system can find relevant chunks; an answer generator and citation checks remain separate. ServiceNow Apriel Voice is similarly a component: it understands speech and can propose tool calls, but its card specifies external synthesis for speech output and an application-managed action loop. Do not grant a voice agent broad tool access because a benchmark looks good.
DiffuSpace Clover tests diffusion-language generation at 35B total / 3B active parameters; its very high speed figures use specific H200/GB300 setups. Liquid's LFM3 is a family rather than a single do-everything file: its 1.7B text model is the easier small local starting point, while other members cover vision, audio and decisions. Baidu's Unlimited OCR 2 tackles long document parsing; the publisher's multi-page instructions and benchmark examples do not remove the need to check page order, tables, handwriting and omissions on your own documents.
The closed comparison: OpenAI's GPT-6.1 Terra is the only non-downloadable model in Matthew's list. OpenAI says it is stronger at visual reasoning and computer tasks than GPT-5.6 Terra, with lower standard API prices. Those are first-party evaluations, and API billing is separate from a ChatGPT subscription. Compare one real task's accepted output, elapsed time and full token bill, not just price per million tokens.
What Should You Try First?
| Your job | First candidate | First check |
|---|---|---|
| A small on-device text assistant | LFM3-1.7B | Fit and quality on five short tasks; confirm the exact quantization and license. |
| Search visual PDFs or mixed documents | Perplexity 0.6B | Recall of relevant pages plus evidence needed to verify them. |
| Generate a small 3D browser scene | CCORE 3D Coder | Runnable code, frame rate and export on the same prompt. |
| Evaluate desktop automation | NorthCUA Nano | A disposable desktop, restricted tools and a human approval gate. |
| Process long scanned documents | Unlimited OCR 2 | Page order, tables, missing text and actual hardware cost. |
| Run a large agent model without local infrastructure | Kimi hosted or Terra API | Use approved data; record accepted outputs, latency and total bill. |
For P-H3-Image or Pika H3 Flex weights in an excluded territory, the first step is license clarification, not an installation tutorial. Hosted-service access can have different terms; verify them with the provider before use.
Copyable Model Evaluation Brief
This is an original test prompt, not supplied by Matthew. It asks for a decision before a large download or account connection.
One task, one model, one baseline
Check access, cost and quality before installation.
Video Chapters
| Time | Topic |
|---|---|
| 00:00 | This week's AI releases |
| 00:37 | Kimi K3.1 Pro weights |
| 02:53 | NorthCUA computer use |
| 04:35 | Pruna P-H3-Image |
| 07:14 | Sarvam PRISM |
| 08:23 | Hydrogen-1.0 science model |
| 10:02 | Perplexity Embed v2 Context |
| 10:51 | Qwen3.5 4B 3D Coder |
| 12:17 | Pika H3 Flex |
| 14:19 | Apriel Voice 27B |
| 15:18 | DiffuSpace Clover |
| 16:20 | Liquid AI LFM3 |
| 17:35 | Baidu Unlimited OCR 2 |
| 18:14 | GPT-6.1 Terra |
Source Note and Official Links
Checked 6 November 2026. The video and supplied transcript are the source for Matthew's demos, reactions and quoted comparisons. Official release pages, model cards and code repositories linked beside each claim establish what publishers say is available; they do not independently verify benchmark rankings, local performance or end-to-end workflow quality. Three important corrections to the narration are that Hydrogen-1.0 is by Deep Principle, Apriel Voice needs external speech synthesis, and both P-H3-Image and Pika H3 Flex have territory-limited model-weight terms.
More primary material: Pruna's technical launch; Sarvam's PRISM explanation; Perplexity's embedding paper; Pika's local inference code; Clover's release notes; and OpenAI API pricing. Model and price pages can change after this date.