The AI log, frontier · open weights · agents and coding
13 Nov 2026
Gemini 4 Hydrogen, Edge0-145B, IBM's LLMON, Brave x402
Google's fast Gemini 4 Hydrogen, an open 145B model for laptops, IBM's structured-prompt Granite previews and pay-per-request search for agents.
By Ali Asif, published
What shipped
- Google DeepMind
Introducing Gemini 4 Hydrogen
A fast execution model priced flat at $1.50 and $7.50 per million tokens across a 1M-token context, scoring 51 on the Artificial Analysis Intelligence Index at default thinking, plus an Orchestrator mode in Antigravity.
- Edge0
Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer
Open weights and an inference framework for a 145B mixture-of-experts model that reports 29GB peak memory and 80+ tokens per second decode on a compatible laptop.
- IBM Granite
LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B
Two experimental preview models trained on a markup that separates instructions from data, which IBM says drives some prompt injection attack success rates close to zero.
- Brave
The Brave Search API now accepts agentic payments via the x402 protocol
Agents can pay per search request in USDC on Base with no account or API key, at $0.005 per request.
Four primary pages I opened and read today: a fast frontier-lab model, an open model sized for laptops, a structured-prompt experiment from IBM and a payment rail for search agents. The numbers below come from those pages, not from my own testing.
Gemini 4 Hydrogen
Google positions Hydrogen as the fast execution model beneath its frontier Gemini 4 Carbon and Gemini 4 Argon, which it says handle deep, long-horizon reasoning. Hydrogen is built for rapid execution across agentic workflows and everyday coding, and it ships to all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Search, plus paid Pro and Ultra subscribers in Antigravity.
Pricing is flat across the full 1 million-token context window: $1.50 per million input tokens and $7.50 per million output tokens. Google reports that Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, against 50 for Argon, and 54 at its highest xhigh setting. It also says the model often finishes tasks faster than others in its class.
The second half of the story is Orchestrator in Google Antigravity, which switches between models such as Carbon and Hydrogen inside one session. On Terminal-Bench 4.0, Google says Orchestrator reached near-Carbon task success while using 24% less quota and finishing 19% faster on average. The idea of a strong model coordinating cheaper worker and verifier agents is the part I would copy in my own workflows.
Edge0-145B
Edge0 open-sourced Edge0-145B and Edge0-120B together with an inference framework and deployment tools. Of 145B total parameters, roughly 10B are active per inference step. The company says each group of 64 experts shares two large matrices, which cuts what has to be stored and moved, and that it trained the model in its on-device architecture from the start rather than compressing a cloud model afterward.
The reported laptop figures are 29GB peak memory, more than 1,100 tokens per second of prefill, more than 80 tokens per second of decode and about 120W typical power, with 128K to 256K context holding roughly the same speed. The post compares coding and reasoning results with specific cloud models, and those comparisons are Edge0's own. It also notes that hardware and electricity still cost money.
IBM's LLMON in Granite 4.1
IBM released granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from Granite 4.1 base models, along with a GitHub toolkit. LLMON is an XML-like markup whose delimiters are reserved special tokens, so structure and content sit in separate parts of the vocabulary. Unique IDs bind each instruction to the data it acts on.
The practical claim is prompt injection resistance: IBM says indirect injections hidden in user data are mostly ignored, with attack success rates near zero on some benchmarks, while general capability stays close to Granite 4.1 Instruct. Both models are labeled experimental previews, and there are converters to move existing JSON into LLMON and back.
Brave Search over x402
Brave added support for the open x402 payment protocol on its Search API. A request without payment gets an HTTP 402 response stating the price, and the agent signs a USDC payment on Base and retries. No account or API key is needed. Brave prices LLM Context and its web, news, image, video and local endpoints at $0.005 per request, and says it covers gas fees for a limited time.
What I'd watch next
Independent runs of Hydrogen against Argon on real coding tasks, and whether Orchestrator's quota savings hold outside Google's benchmark. For Edge0, someone reproducing the 29GB figure on a stock laptop. For LLMON, results from attackers who are not IBM.
Sources
Topics: Google DeepMind, Google, Gemini, IBM, Granite, Antigravity, IBM Granite, ON Device