Tell 1 vs Gemma 4 12B: One You Can Put on a Disk, One You Can Only Phone

The fastest way to understand the difference between Tell 1 and G​emma 4 12B is to try to download both. G​emma 4 12B arrives as roughly 27 GB of BF16 safetensors under Apache 2.0, runs at about 110 tokens per second on a single 24 GB card in a quantised build, and answers to you and nobody else. Tell 1 does not arrive at all. It is Retell's own reasoning model, sitting inside a voice-agent platform that charges $0.07 to $0.31 per minute and does not sell the model, publish a model card for it, or let you point it at your own infrastructure. Everything else in this comparison follows from that asymmetry, including the benchmarks, because one of these models can be measured by anyone with a GPU and the other can only be measured by the company that owns it.

Both are genuinely interesting releases, and they were built for different jobs. G​emma 4 12B is G​oogle DeepMind's encoder-free mid-size open model, released 3 June 2026, designed to do multimodal reasoning on the machine in front of you. Tell 1 is the reasoning layer of a contact-centre stack that handled more than 100 million calls a month before the model existed, and its design goal is the one thing a local model cannot deliver: a first syllable that lands inside a phone caller's tolerance for silence.

Two things called a model, and only one is a file

G​emma 4 12B is 11.95 billion dense parameters with no separate vision or audio encoder — image patches and audio waveforms go straight into the transformer, which is why the family ships as one "unified" checkpoint rather than a stack of towers. It carries a 256K-token context and reads text, images, audio and video frames. The Hugging Face repository has been downloaded more than 14 million times, the chat template was last revised on 9 July 2026, and the licence is Apache 2.0 with no revenue clause, no user-count clause and no separate commercial agreement to negotiate.

Tell 1 has none of those artifacts. There is no repository, no config, no tokenizer, no licence text and no context-window figure published anywhere I could find on Retell's own site — and I looked at the blog, the changelog, the docs index, the newsroom and the company's model-benchmark site. The two places it appears at all are a label on the homepage, where it sits beside "Agent harness" under a section headed "Developed by Retell Labs," and a November update note added to a September changelog about Retell's public voice benchmark. That is the entire public record. It is not a scandal; plenty of platform vendors ship internal models without documenting them. But it does mean that half of this comparison has to be conducted on the vendor's word, and the other half can be conducted by you, tonight, on your own hardware.

The five questions one of them answers in public

Set the two side by side on the dimensions a buyer actually needs, and the shape of the gap becomes obvious. Both sides of each line, because the point is the contrast rather than a score:

• Weights — G​emma 4 12B publishes them, Apache 2.0, self-hostable, fine-tunable, redistributable; Tell 1 publishes nothing and is reachable only as part of a Retell voice agent.

• Size — 11.95B dense parameters, about 27 GB in BF16 and under 7 GB quantised, for G​emma; undisclosed for Tell 1, and the fact that it answers with a 72 ms median inside Retell's harness implies a small, purpose-built model rather than a frontier generalist, but that is inference, not documentation.

• Context — 256K tokens for G​emma 4 12B; unstated for Tell 1, which for a phone agent is less alarming than it sounds, since a call is short and the knowledge base lives outside the model.

• Modalities — G​emma takes text, image, audio and video and returns text; Tell 1 takes the transcript-and-tools payload a voice platform hands it, and what else it accepts is not published.

• Evidence — G​emma has a vendor card (DocVQA 94.9, InfoVQA 88.4, MMLU-Pro 77.2, GPQA Diamond 78.8, all G​oogle-reported) plus an independent Artificial Analysis Intelligence Index of 14 at roughly 110.9 tokens per second; Tell 1 has one row in a benchmark its owner wrote, which is covered below.

Notice which column is longer. This is not a knock on Tell 1's quality; it is a statement about what you can verify before spending money. A model you can download can be tested against your own data in an afternoon. A model you can only rent is tested by renting it.

What Tell 1's headline number actually measures

Retell's vCX Benchmark publishes its results as structured data rather than as a marketing chart, which is more than most vendor leaderboards do. Tell 1 sits in a table of 62 evaluated configurations drawn from 45 models, and its line is strongest at the conversational end: 96.6% on Retell's colloquial dimension and a 72.5 ms median, first in the field on both. The same line shows second place on instruction following and on the row Retell labels hallucination, fifth on tool calling, and a 3.5 cent per minute model cost that puts it 19th of 62. Average the four dimensions and you get 94.45%, which is the figure Retell leads with.

Two caveats belong next to those ranks, and neither is Retell being dishonest — the data simply does not support a cleaner reading. First, Tell 1's row carries benchmark version 17 while 50 of the 62 rows carry version 1; a benchmark that revises its tasks is doing the right thing, but "#1 overall" computed across two different test sets is an aggregate, not a like-for-like result, and the page does not normalise for it. Second, the four dimensions do not agree. Tell 1 is not the best model in its own benchmark at following instructions or at calling tools — GPT-6-Astra scores 89.7% instruction following against Tell 1's 88.6%, and at least one Claude build edges ahead on tool calls. What it wins outright is the two things a caller can hear: it is the quickest and it sounds the most like a person.

One number worth keeping straight, because the two get conflated in coverage: 72 ms is the model's latency inside the harness, not the caller's experience. Retell's homepage quotes a typical agent response time of around 600 ms end to end, which includes speech recognition, orchestration, the model turn and speech synthesis. Both numbers can be true. Only the second one is what a person on the phone perceives.

Where each one belongs in a stack

G​emma 4 12B is a poor fit for a live phone agent, and not because of quality. A 12B dense model on a single GPU is a fine component, but the round trip through your own serving stack, plus speech recognition and synthesis, is what pushes end-to-end response into the range where callers start talking over the agent. G​oogle's own framing for the 12B is agentic work on a laptop — document understanding, screen reading, tool use, coding help — and that framing is accurate. If your job is a hundred thousand invoices a night, a private corpus you cannot send to a third party, or an offline field device, Tell 1 cannot compete, because Tell 1 cannot be deployed there at all.

Tell 1 is the reverse. It cannot leave Retell, cannot be fine-tuned on your transcripts by you, and cannot be benchmarked by a third party without Retell's cooperation. What it can do is answer inside a platform that already owns the telephony, the turn-taking and the voice, and do it in 72 ms, which is the part of the stack that no amount of prompt engineering fixes. If you are buying a contact-centre agent rather than building one, that is the only column that matters.

The middle ground is where most teams actually land: a small self-hosted model for the bulk, deterministic work and a hosted frontier model for the hard turns. Neither of these two models is the natural pick for that second slot, and this is where our own catalogue is relevant. Neither Tell 1 nor G​emma 4 12B is routable through OrcaRouter — we do not host Tell 1, and the G​emma 4 sizes we carry are the 31B Instruct and the 26B A4B, not the 12B. Those two are priced at the provider's list with nothing added, $0.13 input and $0.38 output per million tokens for the 31B and $0.06 and $0.33 for the 26B A4B, which is the honest comparison point if what you wanted from the 12B was "a G​emma 4 I can call without owning a GPU."

Cost, converted into the same unit

Comparing a per-minute platform price against a per-token model price is where most of these write-ups quietly cheat, so here is the conversion rather than a verdict. Tell 1 is not priced separately; it is inside Retell's $0.07 to $0.31 per minute, and the benchmark row that carries it says 3.5 cents per minute of model cost, which sits inside that band rather than replacing it. G​emma 4 12B at Artificial Analysis's blended $0.10 per million input and $0.30 per million output is roughly free per token, and then you pay for the GPU: a 24 GB card running for an hour costs more than a thousand minutes of Retell's cheapest tier, so the crossover only happens at sustained volume with high utilisation.

• Cost unit — dollars per minute of conversation for Tell 1; dollars per million tokens plus GPU-hours for G​emma 4 12B.

• Break-even — G​emma wins on marginal cost the moment your GPUs are already busy; Tell 1 wins on total cost of ownership until the volume is large enough to keep a card saturated.

• Hidden line items — Tell 1 bundles telephony, speech synthesis and the harness into one number; a self-hosted G​emma stack bills you separately for the ASR, the T​TS, the serving layer and the on-call engineer.

Which one to pick

If the requirement is a phone agent that sounds human this quarter, Tell 1 is the only one of the two that is a product, and the benchmark — caveats included — says the thing it is best at is exactly what you are buying. Test it against the model you currently trust inside the same Retell agent, listen to twenty calls, and count the grounding errors, because that is where Retell's own data says it loses.

If the requirement is a model you can inspect, fine-tune, host privately, redistribute inside a product, or run without a network, G​emma 4 12B is not merely the better option — it is the only option on this page. Apache 2.0 and a 256K context at 12B is a combination that gets more useful every month, and the 14 million downloads suggest a lot of teams have already reached that conclusion.

The question worth asking before either purchase is which of those two sentences describes your problem, because they are not the same problem, and no benchmark can tell you which one you have.