Most AI services are wrappers around someone else's API: your request travels to a third-party cloud, and the service controls neither the speed, nor the queue, nor where your data is actually processed. GPTunneL took a different path and built Grom — a family of in-house models that the team researches, trains and runs on its own hardware. This overview covers the whole lineup, the numbers behind it, and why self-hosted inference is a real advantage, not a marketing line.
What "in-house inference" actually means
The full cycle stays inside the team: GPTunneL researches architectures, trains and fine-tunes the models on its own data, and serves them on its own GPU fleet — NVIDIA H200, A100 and A5000 cards, with each workload routed to the hardware that fits it. In practice that gives you three things:
- Data stays on our infrastructure. A request to Grom is processed on GPTunneL's own servers — it is not forwarded to external model vendors' clouds.
- Speed under our control. The queue, priorities and latency are managed by the platform, not by a third party. No "the vendor's API is down, please wait".
- Real SLA for business. Because the hardware is ours, availability and response times can be fixed in a contract — and for strict security requirements, Grom models can be deployed on-premise, inside your company's perimeter.
Grom 1.5 and Grom 1.5 Pro: chat, code, documents
Grom 1.5 is the flagship text model. The numbers: first token in about 500 ms, up to 120 tokens per second at peak, and a 256K-token context window — roughly six hundred pages, enough for a full contract or a day of build logs in a single conversation. The model has vision: drop in a screenshot, a photo of a document or a chart, and it reads the image directly, no separate OCR step.
Grom is tuned for everyday work: rewriting an email, explaining an error in a log, pulling numbers from a screenshot, drafting a plan. And an honest note: it is a mid-size model. Heavy reasoning and large codebases are better handled by frontier models — which sit in the same chat window, one click away.
Grom 1.5 Pro is the senior version: on top of text and image input it can search the web and generate images right in the chat. The card with the current price sits next to this article.
You can start with Grom without topping up your balance — in chat it is free to start. For teams, Grom is available via API on request: the same GPTunneL API key that unlocks the other two hundred models, with agreed rate limits and SLA.
GromGPT: the no-charge model
A separate answer to a popular question — is there a free way in? Yes. GromGPT is a text model in the GPTunneL catalog with no charges at all. It is the simplest way to get to know the platform: sign up and start chatting, no card required.
Grom TV: video with sound in one pass
Grom TV is the video model of the lineup: up to 20 seconds of video with sound in a single generation. Speech, ambient noise and music are synchronized with the frame from the start — not a silent clip you have to dub afterwards.
It takes text, a first-frame photo and an audio track in any combination: bring your own voice recording and the character will act the scene to it, lip-synced; skip the audio and the model voices the scene itself. Resolutions run from 480p for drafts to 720p and 1080p for final renders, with a 4K upscale via Grom Pixel when the clip is headed for a big screen.
Grom TV is strongest in scenes with a single character: avatars, presenters, ads, vertical clips for social media. Complex physics and crowd scenes are not its strong suit yet — but Seedance, Veo and Kling live in the same account.
Grom Art, Nova and Pixel: images
The image wing of the family covers three jobs:
- Grom Art — photo editing by text command: removes objects, outpaints beyond the frame, changes light and style while keeping faces and textures intact.
- Grom Nova — image generation from a description: photorealism, illustration, art.
- Grom Pixel — upscaling up to 100 megapixels: it does not stretch pixels, it reconstructs sharpness and texture — skin, hair, lettering on a sign.
Grom Zvuk, a sound generation model for voice-over, music and effects, is on the way.
Who Grom is for
- Newcomers — GromGPT and Grom 1.5: the first reply arrives in half a second, and you can start without topping up.
- Professionals — Grom 1.5 Pro with web search, plus two hundred models next door: start fast, switch to a frontier model when the task demands it.
- Content creators — Grom TV, Nova and Art in Creative Lab: clips, images and retouching on a single balance.
- Business — API access with the same key as every other model, rate limits and SLA in the contract, on-premise deployment for sensitive data.
The easiest way to try it is the chat: open Grom 1.5, or head to Creative Lab for Grom TV. The full lineup lives on the Grom page.
Grom or frontier models: which to choose
The honest answer is "both". Grom wins on latency, cost and data control — your requests are processed on GPTunneL's own infrastructure. Frontier models — ChatGPT, Claude, Gemini — are still ahead in complex analysis and large code; Midjourney and Flux remain the go-to image tools for many.
That is exactly why a single platform is convenient: in GPTunneL the Grom family sits next to ChatGPT, Claude, Gemini, Midjourney, Kling and the rest. One balance, no subscriptions — route the routine to fast Grom, hand the hard problems to a frontier model, and skip paying for a dozen separate services.
FAQ
Is Grom free? You can start with Grom 1.5 in chat without topping up your balance, and GromGPT is a no-charge model in the catalog. Other models are billed pay-per-use from a single balance.
Where is my data processed? On GPTunneL's own infrastructure — the company's own and leased servers. Requests to Grom are not forwarded to external model vendors. For strict security requirements, on-premise deployment is available.
How fast is Grom 1.5? About 500 ms to the first token and up to 120 tokens per second at peak — it starts typing while larger models are still warming up the request.
Can Grom TV use my own voice-over? Yes. Feed it an audio track and the character will act the scene to it, matching lip movement and gestures. Without a track, the model voices the scene itself.
Do I need to install anything? No. Everything runs in the browser: chat for the text models, Creative Lab for video and images. Sign-up takes a minute.
Bottom line
Grom is what happens when a platform owns the whole stack: a chat model that answers in half a second, a free-to-start entry point, video with sound in one pass, image editing and 100-megapixel upscaling — all served from GPTunneL's own hardware, with your data staying on it. Try Grom 1.5 in GPTunneL: no subscriptions, pay only for what you use — and two hundred more models are waiting on the same balance.