This week’s AI drops came so fast that my to-try list gave up. Thirteen new models landed, and step 2 below is the twist. The creator, Matthew Berman, ran through all of them in one video, and 12 of the 13 are open weights.
What’s new
Here are the ones I’d look at first:
- Kimi K3.1 Pro (Moonshot): A 3 trillion parameter frontier model. It ranks first on Agents Last Exam and second on GDPval, right behind Opus 5.5. You’ll need about 1.2 TB of VRAM, so most people will use an inference provider or wait for quantized versions.
- NorthCUA (Beipo AI): An open computer use family with its own open source harness. It loops through screenshot, think, act.
- P-H3-Image (Pruna AI): A video model rebuilt as an image model. A 4 megapixel image takes 2.6 seconds instead of 12.4, using 8 steps instead of 50.
- Pika H3 Flex: The expert calls it possibly the best text-to-video model he’s used. It can make a 15-second clip with a soundtrack in as little as 7 seconds, starting at a penny per second. A 60-second video costs about 60 cents.
- Hydrogen 1.0: A science model post-trained from GLM 5.3. It beat GPT-6 Astra Max on Frontier Science Research at 25 cents per problem versus 88 cents.
- GPT-6.1 Terra (OpenAI): The one closed model here. Its API price is 50% lower than GPT-5.6 Terra, and it’s reported to match Astra at a tenth of the price per token.
The twist
The biggest model isn’t what surprised me most. It’s how little you lose by going small.
On OSWorld Verified, NorthCUA Pro scores 88 with 397 billion parameters. Mini scores 85 with 35 billion. Nano scores 78.9 with just 9 billion, and the creator thinks it could run on most hardware, maybe even a phone.
That’s about 44 times fewer parameters for roughly 10 fewer points. I keep coming back to that number.
The small-model list keeps going:
- LFM3 (Liquid AI): The 1.7B version fits in about 1 GB at 4-bit, with a 256k token context window.
- Qwen3.5 4B 3D Coder: A 4 billion parameter community fine-tune that writes Three.js scenes with GLB export.
- Clover (DiffuSpace): A diffusion text model with 3 billion active parameters. It finished a computer use demo in 169 seconds, while Qwen 3.6 Flash hit its 30-step limit at 578 seconds without finishing.
Try one in 4 steps
- 🎯 Pick by job, not hype. Choose Sarvam PRISM for fast decisions, Perplexity Embed v2 Context for search across text, images and PDFs, Apriel Voice 27B for voice-driven tasks, or Unlimited OCR 2 for huge scanned documents.
- 📏 Start with the smallest version. Move up only if it fails on your real work.
- 🤖 Let your agent install it. The author’s shortcut is simple: paste the Hugging Face link into your agent and say, “Set this up for me.”
- 🧪 Test it on one real task. Benchmarks won’t tell you if it handles your documents, your screens or your workflow.
Pro tips
- Check hardware before you download. PRISM needs 106 GB of disk space and recommends B200 or H200 chips, which aren’t consumer gear.
- Don’t assume decision models chat. PRISM returns a probability for each answer you supply. It doesn’t generate text.
- Try Terra for big token workloads. It’s reported to finish tasks over 3 times faster on average than Astra and GPT-6.1 Sol.
- Use Unlimited OCR 2 for monster files. It was tested on 5,000-page documents and held up fine.
There’s a lot more in the demos, especially the 3D molecule simulation and Pika’s multi-asset editing. 🎬 Watch the full video to see them in action.