Field Notes

Open multimodal models are maturing. Build real workflows now with a browser agent, OCR, and video.

Today is about open models, local workflows, and safety. You get concrete starting points with no install.

Key AI news

Meta Muse Spark Open lands in Transformers

Muse Spark Open is now in Transformers 5.20. It is multimodal and processes text, images, audio, and video in one run. The model uses a Mixture of Experts. That means it only activates part of the experts per token. The release lists 816B total and 42B active per token. Context length goes up to 1,048,576 tokens. Meta trained it for agentic coding, like multi-file editing, debugging, and tool calls. Checkpoints ship in BF16 and NVFP4. One analyst ranks it the strongest US open weight. That is an assessment, not an independent measurement. For you this means: you can run long, mixed inputs and code help without an API, if you have the right GPUs.12

NVIDIA Nemotron 3.5 Super VL for specialized agents

NVIDIA presents Nemotron 3.5 Super VL as an open, adaptable multimodal model for workflows with seeing, reasoning, and tool use. It is meant to read documents, check sources, and drive multiple tools in one job. It is already available locally in LM Studio. LM Studio lists about 70 GB RAM for the full setup. That limits use to strong workstations. This is relevant if you want sensitive content to stay local.34

Debate over leaked Anthropic bill draft

Reports describe a US AI bill draft pitched by Anthropic. Critics point to strong preemption. That means federal law would displace state law. They also see weaker enforcement than the FRONTIER Act. In parts it would remove powers from states to halt risky models, while federal thresholds would be high. Do not plan compliance only on the basis of single states. This is the cited authors’ reading, not a directed judgment.5

GPT-Rosalind: sponsored access for academic labs

OpenAI offers a Rosalind Academic Plan. Eligible academic labs get sponsored access to GPT-Rosalind and life sciences tools. This targets research, not companies. If you run a university lab, this can lower access barriers.6

MOSS-Image: dense text and layouts in an open image model

A technical report introduces MOSS-Image. The goal is better rendering of dense text and complex layouts in open text-to-image systems. This is research, not a finished product feature. Marketing teams may later get better infographics and posters.7

Highlights for your workday

  • Learn agentic coding in the browser. DeepLearning.AI has a new lab with live preview. You install nothing. Enrollment is free per the provider. Good for teams without a dev setup.8

  • Test video on the cheap. Dreamina offers a basic month for 1.50 US dollars per a partner post. You turn ideas into clips with Seedance. Check price and terms before you buy. This is a provider claim.9

  • Transcribe long PDFs in one go. UnlimitedOcr is included in Transformers 5.20. It processes long documents, can output bounding boxes, and fits mailrooms. You need Python and model setup.1

Tools and updates

  • Transformers 5.20 brings MOSS-Transcribe-Diarize. You get transcript, speaker mapping, and timestamps in one run for long audio. Use cases: meetings, interviews. Requirement: use Transformers, not a finished service.1

  • Nemotron 3.5 Super VL in LM Studio. Vision inputs and tool use local. Expect about 70 GB RAM. Use case: internal review of confidential documents.4

  • LaTeX renderer as a link tool. A lightweight web tool renders LaTeX from the URL fragment. Sharing works by link. Useful for teams with formulas without a heavy editor.10

Try in five minutes

A small agent exercise in the browser with no install. Basis is the new lab from DeepLearning.AI.8

1) Sign up for the lab in the browser. Open a starter project with live preview. 2) Ask the coding agent to build a small function that extracts ticket number, customer, product, and urgency from a support text. Use this text: “Ticket #84219 von Acme GmbH. Produkt: A-12. Problem: Rechnung doppelt. Bitte bis morgen klären. Hoch.” 3) Start the preview. Check if the four fields are filled correctly. Criterion: ticket number 84219, customer Acme GmbH, product A-12, urgency Hoch.

Expected result: a short output with four correctly detected fields. You see the change immediately in the live preview.

A good find

AI:AM discusses how agents get access without losing control. Two points are actionable. Tie security-critical actions like password changes to out-of-band confirmation. Couple LLM code suggestions with formal or automatic checks before anything goes live. This keeps your “control knob” in your hand, even if agents take on routine work.11

Sources

  1. Transformers v5.20.0 (github.com)
  2. Meta shakes up the frontiers, open and closed (interconnects.ai)
  3. NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning (developer.nvidia.com)
  4. @lmstudio: Nemotron 3.5 Super VL from @NVIDIAAI is live in LM Studio!

• 120B total / 12B active MoE • Supports vision input, reasoning • Trained for visual understanding, tool use

Requires (x.com) 5. Leaked Anthropic Draft AI Regulatory Bill Is Unacceptable and Troubling (thezvi.substack.com) 6. Bringing GPT-Rosalind to academic research labs (openai.com) 7. Paper: MOSS-Image Technical Report (huggingface.co) 8. @deeplearningai: 💻Learn by building with an AI coding agent, right in your browser. 👀Live preview, nothing to install.

📚Try https://t.co/lR4uQ0iJ7H’s new agentic coding lab in our course, Building (x.com) 9. @datachaz: 🚨 Dreamina's $1.50 Basic offer is somehow still live.

It gives you a full month to turn saved ideas into actual videos with the Seedance lineup, including Seedance 2.5.

ICYMI the (x.com) 10. LaTeX renderer (simonwillison.net) 11. AI:AM: Where's the Steering Wheel? My Agent Set the Password, Hidden Valence, Proofs Find Linux Bugs (cognitiverevolution.ai)