Our own inference

Grom 1.5

Q: What's under Grom's hood?

A mid-sized model by parameter count: we fine-tune it for the platform's tasks and run it on our own inference. We don't publish the weights or the exact architecture. The practical difference is elsewhere — speed, cost per request and where your text is physically processed are controlled by us, not by an outside vendor.

Q: How is Grom different from GPT-5 or Claude?

Grom starts faster, types faster, costs less to run and lives on our hardware. Frontier models are stronger at hard reasoning, long code and rare languages. That's why they sit right next to it in the same chat: start with Grom, switch when the task calls for it.

Q: Where is the data processed?

On our own infrastructure: our servers plus capacity rented in data centres. A request to Grom is never handed to a third-party model provider.

Q: Can I get API access?

Yes, on request. Grom is opened up with the same GPTunneL API key; we agree TPM and SLA for your load. For sensitive data an on-premise deployment is possible.

The model GPTunneL starts with. It answers first, reads images, holds a long context and runs on our hardware — not in someone else's cloud.

Open in chat The whole Grom lineup

500 ms

to the first token

120

tokens per second at peak

256K

context window

Why

The model that greets you

Grom grew out of a simple problem. Someone signs up, opens the chat — and should get an answer right away, instead of picking between two hundred models and doing math on what the first question will cost. So we tuned our own first model for one thing: a fast, clear answer to an everyday question. Rewrite an email, explain an error in a log, pull the numbers off a screenshot, sketch a plan — the stuff most requests are made of. You can start with Grom without topping up your balance: try it, learn the interface, get a couple of real tasks done. When you hit its ceiling, GPT, Claude, Gemini and two hundred more models are one click away in the same window — and that's where you pay for what you use.

Speed

It starts typing while the others are still thinking

The first token leaves in about half a second, then the stream ramps up to 120 tokens per second. Large models spend a while warming a request up — on a short question that pause is more noticeable than any difference in the answer.

A 500-token answerhatched — waiting for the first token

Grom 1.5our model

4.7 s

0120 tok/s · 500 ms

Claude Sonnet-5

8.8 s

068 tok/s · 1400 ms

GPT 5

9.7 s

062 tok/s · 1600 ms

Qwen 3.5 Plus

11.0 s

055 tok/s · 1900 ms

Our own measurements: a short prompt with no reasoning mode, third-party models called through GPTunneL, Grom on our own inference. Real numbers depend on load, request length and whatever is happening on the vendor's side.

Vision

It sees the picture, not just the text

Drop in a screenshot, a photo of a document or a chart. The model works out what is in the image and answers in text — no separate OCR layer needed.

A boy dribbling a ball on the roadway with a car approaching behind him

What's going on in this photo? Is anything risky?

1200 × 896 · image → ≈1,100 tokens

Screenshots and documents.Text on a photo, a table, a receipt, an error screen — all read straight from the image.
The image joins the shared context.You can keep asking follow-ups without attaching the file again.
Works in chat and over the API.The request format matches every other model on the platform.

Context

256,000 tokens in a single conversation

That's roughly six hundred pages of text. A full contract, a day of build logs, a month of correspondence — put it in the conversation and ask about it without slicing the document up.

256K tokens

A 60-page contract18%
A day of build logs23%
Conversation history12%
Left for answers and follow-ups47%

Prefix cache in GPU memory

A long document isn't recomputed for every question: the attention keys and values already calculated stay in the accelerator's memory. The first question about a file pays for processing, the next ones start from an already warm context.

First question about the document1900 ms

Next question, context cached480 ms

Comparison

Where Grom wins — and where it doesn't

Honestly: this is a mid-sized model. It wins on speed, price and the fact that it runs on our own machines. Hard analysis and large codebases are better handed to frontier models — they're in the same chat window, one click away.

	Grom 1.5our inference	Claude Sonnet-5anthropic	GPT 5openai	Qwen 3.5 Plusalibaba
First token	~500 ms	1.3–1.8 s	1.5–2 s	1.8–2.5 s
Generation speed	up to 120 tok/s	~70 tok/s	~60 tok/s	~55 tok/s
Context window	256K	1M	400K	1M
Images as input	yes	yes	yes	yes
Hard reasoning, large code	basic level	clearly stronger	clearly stronger	stronger
Price in chat	start with no top-up	per price list	per price list	per price list
Where the request runs	our servers	vendor cloud	vendor cloud	vendor cloud

Speed figures are our own measurements inside GPTunneL on a short prompt with no reasoning mode; for frontier models they swing a lot with vendor load. Context windows are per provider documentation as of July 2026.

Infrastructure

A heavy request doesn't slow down a light one

There's a balancer in front of the model. It looks at the weight of a request — how many input tokens, whether a long context is needed — and decides which hardware should run it.

Short questionup to 2K tokens

Conversation with a file20–60K tokens

Long contextup to 256K tokens

Balancerweighs the request and holds the queue

pool AA5000 24 GBshort requests, quantised model

pool BA100conversations with files, mid-sized context

pool CH200long context and large batches

Part of the fleet is our own servers, part is rented in data centres. Grom is a mid-sized model: quantised, it fits on relatively modest cards, so we keep it warm across several pools at once and ride out peaks without a queue.

API

Grom over the API — on request

The model doesn't only live in chat. For teams we open API access with clear limits and commitments.

The same key

Grom is called with the same GPTunneL API key as every other model on the platform. No separate integration.

TPM for your load

We agree a tokens-per-minute limit and hold capacity for you instead of a shared queue.

SLA in the contract

Availability, response time and processing priority are put in writing.

On-premise

If the data can't leave your perimeter, we deploy Grom inside it, on your own hardware.

Need Grom over the API?

Write to us — we'll go through your load, limits, SLA and pricing.

Message us on Telegram Business solutions

Questions about Grom

/01Is Grom paid?

You can start with Grom without topping up your balance — that's exactly what it was built for: a new user should get a working model, not a payment form. Every other model is billed per the price list, and you only pay for what you use.

/02What's under Grom's hood?

/03How is Grom different from GPT-5 or Claude?

/04Where is the data processed?

/05Can I get API access?

Try Grom right now

Signing up takes a minute, the first answer takes half a second. The whole Grom lineup and two hundred more models are right there.

Get started The Grom lineup