Google's Antigravity team (Sachin Kotwani, Taylor Mullen) announced first-class local-model support in the Antigravity SDK, claiming vendor agent workflows can now run on local hardware — a major vendor agent SDK converging on local execution.
state: corroboratedheat: mediumuncertainty: mediumconvergesscott: highagent-harnesses local-inference antigravity-sdkGoogleSachin KotwaniTaylor Mullen
Surfaced 2026-09-25T18:02:02Z — Google announces 'Support for Local AI Models in the Antigravity SDK' on the Google Developers Blog. — Announcement window closed: at ~75h the thread is at zero velocity (current peer percentile 32.5; the four velocity spikes were the initial pickup, already priced in) and no new evidence, implementations, or independent verification has arrived — the magnitude-valve flag reflects cumulative Reddit score on a single discussion thread plus its source blog, not an expanding periphery. Formalizing corroborated on the independent third-party coverage of the self-hosting direction (not the still-unverified performance claims) and cooling to a quiet watch for hands-on offline-execution reports.
What is this?
On Sept 23, 2026, Google's Antigravity team (Sachin Kotwani, GPM; Taylor Mullen, Principal Engineer) announced on the Google Developers Blog that the Antigravity SDK now supports local workflows: agentic loops (file reads, code generation, terminal execution, testing) running entirely offline on local GPU/RAM, with initial support for Gemma 4 26B A4B via Google AI Edge's LiteRT. Antigravity is Google's agent-first development platform (IDE, CLI, SDK, Managed Agents, Enterprise); the SDK exposes the same harness that powers Google's own products, and since Antigravity 2.0 (I/O 2026) developers can already host the agent loop on their own infrastructure. Most of the evidence is vendor-authored; the local-execution performance claims are Google's own and unverified in the supplied material, though third-party coverage (dev.to, apidog, Gadget Hacks) corroborates the SDK's self-hosting direction while noting the dependency-gravity concern — the open blueprints still route through Google surfaces.
Why it matters to Scott
A major vendor agent SDK (Google, via the same harness powering its own products) has newly arrived at what Scott has built and argued privately for years — local/offline agent loops on his own hardware, from the 2023 GPT4All experiments through today's Ollama endpoint on gamepc and the --ollama route in his ask CLI. This is a dated-receipts convergence: Google validating the local-execution leg of his Model Barbell and Scout–Senior Split, with the same model (Gemma 4 26B) the radar already tracks running locally on a Mac mini. The catch his vendor-lock-in and Sovereign Software Assurance concepts predict is confirmed by third-party coverage — the open blueprints still route through Google surfaces — which turns this from 'the world agreeing with him' into a testable claim (does the harness work when the Google surfaces are unreachable?) and a concrete build: wire Antigravity's local path to gamepc and see whose dependency-gravity wins. Evidence is vendor-authored and the performance claims unverified, so this warrants a case, not a verdict.
ip:concept.model-barbellip:framework.scout-senior-splitip:concept.vendor-lock-inip:framework.sovereign-software-assurancedev:project.gamepcdev:technology.ollamaradar:concept.local-inferenceradar:concept.agent-harnessesradar:concept.coding-agentsradar:gemma4-m4-ollama-wired-memory
queries asked of Scott's wikis
- agent harness local inference offline execution
- local models for coding agents small model agent loop
- Gemma on-device LiteRT local agent
- vendor SDK self-hosting agent runtime data residency
- open-weight models vs hosted API agent workflows
- agent memory and RAG running on local hardware
Measured heat
now 0 pts/hpeak 1 pts/hcomments 0/hpeers p16momentum: steady3 platformsage 458h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
How the heat travelled
pace: p81 vs 1032 stories at the 336h mark (now 458h old) — ahead of fan-coding-harness-component-study (1.0x), behind system76-thelio-mira-ai (1.0x)
Evidence (3) — ⭐ canonical anchor
| source | object | author | score | comments |
|---|
| 🟠 reddit | Introducing Support for Local AI Models in the Antigravity SDK LocalLLaMA Retrieved article excerptOpen article · Retrieved 2026-09-23T21:46:39.619108+00:00 [Google for Developers](https://developers.google.com/)
# Introducing Support for Local AI Models in the Antigravity SDK
SEPT. 23, 2026
[Sachin Kotwani](https://developers.googleblog.com/search/?author=Sachin+Kotwani)
Group Product Manager
[Taylor Mullen](https://developers.googleblog.com/search/?author=Taylor+Mullen)
Principal Engineer
Share
- [Facebook](https://www.facebook.com/sharer/sharer.php?u={url} "Share on Facebook")
- [Twitter](https://twitter.com/intent/tweet?text={url} "Share on Twitter")
- [LinkedIn](https://www.linkedin.com/shareArticle?url={url}&mini=true "Share on LinkedIn")
- Mail
agy\_litrt
posted in:
- [AI](https://developers.googleblog.com/search/?technology_categories=AI)
- [Announcements](https://developers.googleblog.com/search/?content_type_categories=Announcements)
- [Explore](https://developers.googleblog.com/search/?content_type_categories=Explore)
Previous
Next
Related Posts
[Why client SDK generation belongs in the open
AI
Announcements
Learn
Why client SDK generation belongs in the open
SEPT. 17, 2026](https://developers.googleblog.com/why-client-sdk-generation-belongs-in-the-open/)
[Colab is now part of your Google AI plan
AI
Announcements
Events
Colab is now part of your Google AI plan
SEPT. 22, 2026](https://developers.googleblog.com/colab-is-now-part-of-your-google-ai-plan/) | dryadofelysium | 320 | 43 |
| 🟧 echo.blog ⭐ | Google announces 'Support for Local AI Models in the Antigravity SDK' on the Google Developers Blog. | Sachin Kotwani and Taylor Mullen, Google | — | — |
| 🟧 hn | Antigravity Excel Engine – Dual-Core Excel Automation for AI Agents | fgrandon | 3 | 0 |
Interpretation history
2026-10-10T04:59:54Z
Case has been quiet for 17+ days since the Sept 26 cooling decision. The new HN attachment (Antigravity Excel Engine, 3 pts/0 comments) is negligible ecosystem noise — not an expanding periphery, independent verification, or hands-on offline test. Measured heat confirms near-zero velocity (0.5 pts/h, 0 comments/h, peer percentile 46.8). Heat now aligns with the 'quiet watch' posture established Sept 26.
2026-10-10T04:39:54Z
evidence attached: hn.story.50028857 — Agent-focused Excel automation tool named for Antigravity SDK; concrete ecosystem tooling for agent-to-office-app workflows
2026-09-25T17:54:45Z
magnitude valve eligible (multi-platform, top-decile engagement) and never alerted; deterministic escalation to deliver
2026-09-23T21:54:41Z
grounded: converges/high — A major vendor agent SDK (Google, via the same harness powering its own products) has newly arrived at what Scott has built and argued privately for years — loc
2026-09-23T21:48:56Z
case created — First-party Google announcement of local-model support in its agent SDK, directly on the local-inference/agent-harness intersection with strong community pickup.
Decision trace
- 10-10 17:04attention_routeThe editor compared this story and chose to keep watching.
- 10-10 15:59repriceCase has been quiet for 17+ days since the Sept 26 cooling decision. The new HN attachment (Antigravity Excel Engine, 3 pts/0 comments) is negligible ecosystem noise — not an expanding periphery, inde
- 10-10 15:40attention_routeFirst-party vendor SDK convergence on local execution validates Scott's dated-receipts architectures (Model Barbell, Scout–Senior Split, Sovereign Software Assurance). The catch — open blueprints
- 10-10 15:39attention_candidateattach
- 10-10 15:39attachAgent-focused Excel automation tool named for Antigravity SDK; concrete ecosystem tooling for agent-to-office-app workflows
- 10-10 15:39propose_attachAgent-focused Excel automation tool named for Antigravity SDK; concrete ecosystem tooling for agent-to-office-app workflows
- 10-06 21:42review_dormantscheduled targets exhausted or 28 quiet days
- 10-06 21:42drop_targetsquiet through full ladder or over cap 8
- 09-26 04:02pushGoogle announces 'Support for Local AI Models in the Antigravity SDK' on the Google Developers Blog. — Announcement window closed: at ~75h the thread is at zero velocity (current peer percen
- 09-26 03:54repriceAnnouncement window closed: at ~75h the thread is at zero velocity (current peer percentile 32.5; the four velocity spikes were the initial pickup, already priced in) and no new evidence, implementati
- 09-26 03:54alert_heldGoogle announces 'Support for Local AI Models in the Antigravity SDK' on the Google Developers Blog. — Announcement window closed: at ~75h the thread is at zero velocity (current peer percen
- 09-26 03:54alert_routeGoogle announces 'Support for Local AI Models in the Antigravity SDK' on the Google Developers Blog. — Announcement window closed: at ~75h the thread is at zero velocity (current peer percen
- 09-25 11:21sensor_dirtyvelocity_spike
- 09-25 01:21sensor_dirtyvelocity_spike
- 09-24 17:21sensor_dirtyvelocity_spike
- 09-24 10:21sensor_dirtyvelocity_spike
- 09-24 07:54groundA major vendor agent SDK (Google, via the same harness powering its own products) has newly arrived at what Scott has built and argued privately for years — local/offline agent loops on his own hardwa
- 09-24 07:48createFirst-party Google announcement of local-model support in its agent SDK, directly on the local-inference/agent-harness intersection with strong community pickup.