Run open agents locally, use zero-shot forecasts, delegate agents safely

Important AI news

NVIDIA ships Nemotron 3.5 Super VL and Bolt from server to edge

NVIDIA releases Nemotron 3.5 Super VL with 120 billion parameters and open weights. Plus two edge models: Bolt 74B-A9B, which runs locally in 24 GB VRAM, and Bolt E4B for 4 GB, announced for October 23. Super VL targets agent tasks with text and image like tool use, multi step workflows, OCR, and tables in PDFs. NVIDIA claims 5x higher throughput versus models in the same class on the Artificial Analysis Intelligence Index and top scores on PinchBench and BrowseComp. Those are vendor claims. Super VL and Bolt 74B-A9B are available, Bolt E4B to follow. What this means for you: You can run agents locally on RTX GPUs with 24 GB, keep sensitive documents in house, and continue training Super VL on your own data.¹

Mistral releases Timestral for zero shot forecasting

Timestral is a family of time series base models, from 5M to 209M parameters. Zero shot forecasting means predictions without project specific training on your data. According to the model card, Timestral-209M reaches on GIFT-Eval MASE 0.658 and CRPS 0.446, on TIME MASE 0.626 and CRPS 0.522. Timestral-28M is at MASE 0.678 and CRPS 0.461 on GIFT-Eval. All figures come from the named leaderboards. The model takes up to 32,768 steps of context and outputs 21 quantiles. Inference runs on CPU or GPU, license is a modified MIT. You can quickly test baselines for demand, inventory, or energy use before you fine tune.²³⁴⁵

Microsoft Research shows Magnetar for provable execution

Magnetar adds short cryptographic proofs to program outputs. The proofs make correctness verifiable in milliseconds. A zero knowledge proof is a proof that confirms correctness without revealing secret inputs. Magnetar proves programs in Rust, Go, and C and is on average 10 to 39 times faster than comparable systems according to Microsoft. The security rests on standard hash functions. This is a research release, not a product. Why it matters: Agent actions could ship with machine verifiable proofs in the future. That would raise trust in high criticality workflows.⁶

Highlights for your workday

• More autonomy with control: AWS sketches Trust in Depth for agents. Four questions structure governance: set scope, know identity, limit at runtime, review after. Where rules are precise, automated closure helps. That means mathematical checking of rules across all cases. You get a strong start if you write rules for one agent job, minimize access, and route real ambiguity to humans.⁷

• Review contracts faster: Sullivan & Cromwell built the Agreement Analyzer with OpenAI, which flags issues in draft contracts. The service is not a generally available product. You can reuse the pattern: define a fixed list of issues, auto label clauses, and escalate only disputed passages. That relieves lawyers without giving up control.⁸

• Plan big peaks like Meta: For the 2026 World Cup they used capacity buffers, country specific degradation rules, and early warning signals to carry the load spikes. Meta even skipped a freeze and kept releases going. Map that to your events: hold buffers, define degradable secondary load per market, prevent cascades, and set alert thresholds.⁹

Tools and updates

• Muse Spark Open: Meta presents a coding optimized MoE model with 816B experts and 42B active parameters with a 1M context window. The goal is agentic coding workflows with tool calls and codebase understanding. It requires NVIDIA GB300 NVL72 infrastructure.¹⁰

• MLX on the Mac: Apple’s MLX enables local and distributed inference on Mac Studio with Apple Silicon. Good for privately testing new inference and fine tuning methods if you already develop on macOS.¹¹

• Open learning: NVIDIA’s Deep Learning Institute launches a freely usable open source course hub for accelerated computing and deep learning. No signup, no fees. Good for internal upskilling tracks.¹²

Try in five minutes

Make agents safer: a mini check before launch.

1) Pick one agent job with clear value, like answering invoice queries. Write down purpose, allowed data and tools, and three things the agent must not do. 2) Write one precise rule that a machine can check. For example: payments over 5,000 euros only after manager approval. Decide what happens if unclear, like abort with a follow up question. 3) Define runtime limits: maximum access, allowed systems, stop conditions. Set a simple run log: time, inputs, actions, result, rationale, reviewer.

Expected outcome: a one page agent brief with binding rules. Check: A colleague can safely decide from it what the agent may do alone and what not.⁷

Sources

1. Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models (blogs.nvidia.com)
2. mistralai releases mistralai/Timestral-209M on Hugging Face (huggingface.co)
3. mistralai releases mistralai/Timestral-28M on Hugging Face (huggingface.co)
4. mistralai releases mistralai/Timestral-123M on Hugging Face (huggingface.co)
5. mistralai releases mistralai/Timestral-5M on Hugging Face (huggingface.co)
6. Proof-based general-purpose computing with Magnetar (microsoft.com)
7. Trust in Depth: Giving agents more autonomy without losing control (aws.amazon.com)
8. Sullivan & Cromwell brings legal expertise into agreement review (openai.com)
9. 2026 World Cup: Breaking Records, Spiky Charts and Reliability at Scale (engineering.fb.com)
10. Scale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72 (developer.nvidia.com)
11. Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio (machinelearning.apple.com)
12. NVIDIA Expands Advanced Technology Access for Europe’s Scientific Community Through Open Learning Resources, Collaborations and Compute (blogs.nvidia.com)