Skip to content
Agentdex

Model Pulse #19 · Thursday, Nov 19, 2026

Google releases Gemini 4 Neon, Tencent opens Hy4 Flash, Netlify relaunches for agents, and Palo Alto Networks buys Applied Compute for $4.5B

The seven AI stories since yesterday's issue that matter most for AI coders and indie game devs. Each one links to its primary source.

Sponsored Sponsor this issue: reach AI coders and solo founders in Model Pulse.

Nov 18, 2026

DEALS: Palo Alto Networks agrees to buy Applied Compute for $4.5B to bring secure open-weight model training to enterprises, and Linq raises $90M led by a16z for agent messaging

Palo Alto Networks announced on Nov 18 that it will acquire Applied Compute in an all-stock deal valued at $4.5B, subject to adjustments. Applied Compute runs a platform for post-training and serving open-weight AI models, so a company can tune a model on its own data and run it where it wants. Palo Alto says it believes enterprise AI will be hybrid, with a large share of companies running open models tuned to their own data, cost and control needs, and that it wants to be the partner that delivers those models securely, in the public cloud or in sovereign environments. Applied Compute will keep serving AI-native companies, and its research team will help train Palo Alto's own large and small models. Co-founder and CEO Yash Patil becomes Palo Alto's Co-Chief AI Officer after the deal closes. The 8-K filing prices the shares at $337.16 each, the 15-day average, and the deal is expected to close in Palo Alto's fiscal third quarter of 2027, pending regulatory clearance. Reuters reports that Palo Alto shares rose about 9% after hours as it also raised its full-year revenue forecast. Also on Nov 18, Linq announced $90M in new funding led by Andreessen Horowitz, with Ben Horowitz joining its board and total funding now above $120M. Linq is the infrastructure that lets developers build agents people can text or call. DoorDash, Town and hundreds of others use it, and it has processed more than 1 billion messages. It is now an Apple Messaging Service Provider, and it is adding HD Calling for agents and access to real computers so agents can use software, with browser-agent company Kernel among the first customers. Linq did not disclose a valuation.

Source: https://www.paloaltonetworks.com/company/press/2026/palo-alto-networks-to-acquire-applied-compute--bringing-secure-open-weight-ai-to-the-enterprise

Nov 18, 2026

MODELS: Google releases Gemini 4 Neon, which tops the Artificial Analysis Coding Agent Index at $3/$15 per million tokens and can plan for Hydrogen subagents in Antigravity

Google released Gemini 4 Neon on Nov 18, a model that sits between Gemini 4 Hydrogen and Gemini 4 Carbon. It costs $3 per million input tokens ($0.15 cached) and $15 per million output, with a 1-million-token context window and four thinking levels from Low to Extra High. It is rolling out in the Gemini API in Google AI Studio, Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra. Google says Neon ranks first on the Artificial Analysis Coding Agent Index (58.6), third overall on the Artificial Analysis Intelligence Index (61 at Extra High), and scores 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default High setting. On the SWE-sweep benchmark, it scores 12.8% in the mini-SWE harness and 39.0% inside Antigravity (company figures). Google also says Neon makes fewer redundant file reads, handles context compaction in long sessions, and had its lowest fabrication rate yet in internal Antigravity testing. The Antigravity team says Neon at High costs 32% less per task than Carbon, and that Orchestrator can now pair Neon as the planner with Hydrogen as the executor in one session, handing work back to Neon when Hydrogen gets stuck. Its demo has Neon with /swarm scaffolding a multi-level two-player game that runs in the built-in browser while the agent plays as the second player. 9to5Google reports that Neon replaces Gemini 4 Argon in the Gemini app's model picker.

Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/

Nov 17-18, 2026

OPEN WEIGHTS: Tencent puts Hy4 Flash on Hugging Face, a 220B mixture-of-experts model with 12B active parameters, a 1M-token context and image input

Tencent published the weights of Hy4 Flash on Hugging Face, along with an FP8 version and a pre-trained base model, and also on ModelScope and GitCode. Tencent calls it its most cost-effective Hy model so far. The model card lists 220B total parameters, including 64B of n-gram embedding parameters (Engram), and about 12B active per token. It reads text and images, writes text and has a 1-million-token context. It mixes sparse and sliding-window attention, routes 16 of 512 experts per token, and includes a 0.85B built-in drafter (DFly2) for speculative decoding. Tencent says that, combined with megakernels and 4-bit quantization, this pushes decoding past 150 tokens per second (company figures). Its own charts compare cost and score across seven public agent benchmarks, including DeepSWE, Terminal-Bench 4.0 and OSWorld 2.1, plus WorkIndex, an internal benchmark built from tasks in Tencent's WorkBuddy app. There are official vLLM and SGLang images. ModelFit, reading the Hugging Face files, notes that the weights went up on Nov 17, that the BF16 repo is about 441 GB and the FP8 repo about 287 GB, that an official 4-bit GGUF is about 157 GiB, and that llama.cpp support was still an open pull request. The license is the Tencent Hy License, which is based on Apache 2.0 but is not Apache 2.0. Running the model yourself is fine, but a hosted model service needs a separate license from Tencent once it passes $50M in yearly revenue or 100 million monthly users. Read the LICENSE before you build a hosted product on it.

Source: https://huggingface.co/tencent/Hy4-flash

Nov 18, 2026

CODING TOOLS: Netlify relaunches as an AI app platform, where you prompt an agent or push to Git into the same app, and says it is nearing $130M ARR

Netlify rolled out a redesigned product and a new brand on Nov 18. In the new experience you write or dictate a prompt, pick an agent and a model, and watch the app build in a live preview next to the conversation. You can still drag and drop files, import from Git, connect the agent you already use, or run netlify deploy. All of those land in the same app, so a developer pushing to Git and a teammate prompting an agent are building the same thing. Every change creates a version with a preview you can review, archive or publish, even without a Git repo, and you can password-protect an app or require a team login in a few clicks. Databases, logins, forms and functions are built in. Netlify's press release adds Agent Runners for building with supported agents, Netlify Server for real-time apps, persistent connections and longer tasks such as AI agents (more than 10,000 servers already run on it, the company says), and more than 150 connectors such as Notion, Stripe, Airtable and Linear that let agents build from your existing data. Connectors arrived on Nov 17 on credit-based plans. Netlify expects about $130M in annual recurring revenue by the end of 2026, nearly 130% growth year on year, and CEO Matt Biilmann writes that it onboarded more than 100,000 new users in one day (company figures). The new experience is rolling out to most new users first, and existing users can join a waitlist.

Source: https://www.netlify.com/changelog/2026-11-18-new-netlify-experience/

Nov 18, 2026

AGENTS: Google Labs opens a waitlist for Stitch Loop, a 24/7 product agent that watches your live metrics and opens pull requests you must approve

Google Labs introduced Stitch Loop on Nov 18, a long-running product agent that picks up where coding agents stop: after the pull request merges. You connect GitHub and your product context, such as logs, user behaviour and feedback, through more than 180 integrations, custom MCP servers or local CLI uploads. Stitch Loop then runs ongoing "loops" toward goals you set. Each loop suggests an improvement with evidence, offers one or more solutions with tradeoffs (and design options where relevant), writes the code in an isolated cloud sandbox, and opens a GitHub pull request, with screenshots and videos where useful. It then watches live data to learn from what worked and keeps watching once the goal is met. Nothing merges until you approve it. Three loops come built in, for security, latency and accessibility, and you can add your own, such as checkout conversion. The security loop pairs Gemini 4 Carbon with CodeMender, Google DeepMind's code-security agent. Google says Stitch Loop can work as a service layer for coding agents such as Antigravity, or inside Slack. It is a Labs experiment. The waitlist at stitch.google.com/loop/waitlist is open worldwide, with priority for eligible Google AI Ultra subscribers, and no price or access date has been announced. For a small team or a solo game dev, the useful idea is a bot that keeps fixing what you shipped. The cost is giving an agent access to your repo and your metrics, so decide which ones before you connect.

Source: https://blog.google/innovation-and-ai/models-and-research/google-labs/stitch-loop/

Nov 18, 2026

OPEN MODELS: NVIDIA post-trained Nemotron with Factory's Droid harness inside its own air-gapped network, and reports big Terminal-Bench gains

Factory published a case study on Nov 18 showing how NVIDIA ran Factory's agent platform on its own premises, in a fully air-gapped network, to post-train its open Nemotron models. Researchers ran long agent tasks, read full execution traces and worked with Factory to find specific failure modes, such as hallucinated tool calls, wrong argument formats and poor error recovery. They then used the harness to reproduce failures and generate examples of correct tool use. For Nemotron 3 Ultra, NVIDIA trained on 20,000 rows of data from the Factory harness, using a reinforcement-learning method called PivotRL. NVIDIA reports that after multi-harness training that included this data, Nemotron 3 Ultra rose from 38.5 to 49 on Terminal-Bench 2.0, and a preliminary Nemotron 3.5 Super VL checkpoint rose from 52.11 to 60.4 on Terminal-Bench 2.1. These are NVIDIA's own evaluations, and the gains come from a wider pipeline, not from Factory data alone. NVIDIA researcher Junkeun Yi says the improvements carried over beyond Droid to general agentic instruction following. Jian Zhang, who directs Nemotron post-training at NVIDIA, says Droid was the only commercially supported harness NVIDIA deployed in its air-gapped environment. For AI coders, this suggests open models are getting better as agents in harnesses beyond the ones they were trained in. For teams with strict data rules, it shows a coding agent and an open model can both run behind your own firewall.

Source: https://factory.com/case-studies/nvidia

Nov 17-18, 2026

GAME DEV: Roblox will convert selected Unity, Unreal and other shipped games to Roblox for free with an AI-assisted pipeline, and you keep the IP

Roblox opened applications for a programme that converts existing games to Roblox at no cost. The programme is for developers whose games are already released on the App Store, Google Play, Steam or a console store, who hold the rights and have the full source code and assets. Roblox says it is building AI-assisted tools to convert a game's source code and assets into a Roblox version, run by its own team, working directly with the developer. Selected teams sign an agreement, then share their code and assets, leaving out any third-party parts they can't license. Roblox says developers keep ownership of their game and IP and own any Roblox version delivered to them, that nothing is published unless they choose to, that no money changes hands either way, and that submitted code and assets will not be used to train or fine-tune AI models. It is starting with a small number of games across genres, and warns that a Roblox version isn't guaranteed because some games may not convert successfully. Hitmarker reports that games from any engine, including Unity and Unreal, can apply. Roblox's pitch to indie devs is reach: it cites 138 million daily active users and $1.6B paid to creators in the last 12 months, and says one Roblox version covers phone, PC and console, with servers, matchmaking and payments handled for you. Before you apply, check which plugins and assets in your game you can actually license.

Source: https://create.roblox.com/docs/creator-programs/convert

Ship faster this week

Grab the free Agentdex starter kit with prompts and agent configs for Cursor, Claude Code and Codex.

Get the free kit

Model Pulse

Daily AI model drops, tools, and what they mean for AI coders.

We store your email with Resend and use it only for Model Pulse. See privacy. Missed the last issue? Read Model Pulse #18.