LLM Weekly — W35: Nvidia reportedly buys Hugging Face for $12.9B, IBM puts reasoning inside the open weights

Open weights had a very good week. Open distribution had a very bad one.

Nvidia agrees to buy Hugging Face for $12.9 billion

The Information reported that Nvidia has agreed to acquire the open-source model repository — the de facto neutral commons where open-weight models, datasets, and inference demos are hosted and pulled from. Terms and timing beyond the headline number aren't established, and this is single-outlet reporting with no statement from either company in this week's flow. If it closes, every "we'll self-host to avoid vendor lock-in" plan at every enterprise runs through infrastructure owned by the company selling the compute those models run on. Reported by The Information.

IBM's Granite 4.2 moves reasoning into the base model

The updated open enterprise family trains reasoning into the weights rather than bolting it on at the prompt layer, and adds reinforcement learning aimed specifically at multi-step tool use. The relevant part isn't the benchmark line — it's that the self-hostable tier just picked up the capability that was the main argument for sending regulated data to a frontier API. Banks, health systems, and government buyers now have a reasoning-grade option inside their own perimeter. Via MarkTechPost.

Four model launches in one cycle — all capability claims vendor-stated

DeepSeek Vision, Ornith 1.5, GEN 1.5, and SenseNova U1.5 all shipped inside the same week, several from Chinese labs, spanning vision, general reasoning, and generation. None of them has independent benchmark verification yet. Worth noting the release cadence itself: the gap between frontier announcements and open-weight follow-ons keeps compressing, and it's compressing fastest outside the US labs. Roundup via AI Search.

Perplexity's Portable Computer enforces the agent sandbox in the OS

Perplexity shipped an agent runtime that executes on a desk-side DGX Spark box, with the sandbox enforced by the operating system rather than by instructions in the prompt, and no metered token cost for steps that run locally. That's a direct answer to the two objections that stall enterprise agent rollouts: unbounded per-token spend, and the agent talking itself out of its own guardrails. It's also a quiet admission that prompt-level constraints don't hold under pressure. Via MarkTechPost.

Evoke open-sources a world model that remembers

Evoke generates interactive environments and holds state across a session instead of regenerating from scratch each frame. Games are the obvious read; the more interesting one is agent training. Persistent, generated environments are the substrate simulation teams have been paying proprietary vendors for, and it just went free. Via AI Search.

What to watch

Two threads worth tracking into W36. First, whether anyone independently benchmarks last week's release wave — four launches with zero third-party numbers is a pattern, not a coincidence. Second, whether the Hugging Face deal draws a mirroring response: if the commons gets an owner, the hedge is a second commons, and the labs with the most to lose from a Nvidia-run repository are the ones with the resources to build one. Open licence, closed distribution is a stable arrangement right up until it isn't.

---

This post is also published on our Substack newsletter at edge-ai.forum. Subscribe for the weekly roundup direct to your inbox — fresh AI news, executive context, and devices + robotics every Friday morning.