Jamba TPS calculator

Open weights AI21 Labs 51.6B parameters March 2024

Each card below is assessed against this model at the context length and minimum quality you choose. Speed is an estimate for a single request, calculated from the card's memory bandwidth and the size of the model once compressed.

Calculated for this model

93 of 818 cards that can run it

Smallest card that fits

Radeon PRO V710

28 GB · Q3_K_M · 37.4 tok/s

Fastest card

B200

282 tok/s · 180 GB

Which GPUs can run Jamba?

Set the inputs, read the answer

Context length tokens

A longer conversation needs more memory, which can push this model off smaller cards.

Minimum quality

Hides cards that would only fit the model by compressing it below this point.

Search cards

Card memory

Also show cards that cannot run it

93 cards match

Calculating

					Needs	Quantisation	Fit
282 tok/s 169–452 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	55.0 GB	Q8_0	Comfortable
282 tok/s 169–452 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	55.0 GB	Q8_0	Comfortable
225 tok/s 135–361 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	55.0 GB	Q8_0	Comfortable
225 tok/s 135–361 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	55.0 GB	Q8_0	Comfortable
180 tok/s 108–289 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	55.0 GB	Q8_0	Comfortable
173 tok/s 104–276 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	55.0 GB	Q8_0	Comfortable
173 tok/s 104–276 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	55.0 GB	Q8_0	Comfortable
165 tok/s 99–264 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	55.0 GB	Q8_0	Comfortable
162 tok/s 97–259 · low confidence	DRIVE A100 PROD NVIDIA	32 GB	1,870 GB/s	May 2020	28.0 GB	IQ4_XS	Tight
162 tok/s 97–259 · low confidence	GRID A100A NVIDIA	32 GB	1,870 GB/s	May 2020	28.0 GB	IQ4_XS	Tight
155 tok/s 93–248 · low confidence	GeForce RTX 5090 NVIDIA	32 GB	1,790 GB/s	Jan 2025	28.0 GB	IQ4_XS	Tight
155 tok/s 93–248 · low confidence	GeForce RTX 5090 D NVIDIA	32 GB	1,790 GB/s	Jan 2025	28.0 GB	IQ4_XS	Tight
147 tok/s 88–235 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	55.0 GB	Q8_0	Comfortable
147 tok/s 88–235 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	55.0 GB	Q8_0	Comfortable
147 tok/s 88–235 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	55.0 GB	Q8_0	Comfortable
139 tok/s 83–222 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	55.0 GB	Q8_0	Comfortable
127 tok/s 76–203 · low confidence	A100 PCIe 40 GB NVIDIA	40 GB	1,560 GB/s	Jun 2020	31.0 GB	Q4_K_M	Tight
127 tok/s 76–203 · low confidence	A100 SXM4 40 GB NVIDIA	40 GB	1,560 GB/s	May 2020	31.0 GB	Q4_K_M	Tight
127 tok/s 76–203 · low confidence	A800 PCIe 40 GB NVIDIA	40 GB	1,560 GB/s	Nov 2022	31.0 GB	Q4_K_M	Tight
119 tok/s 71–190 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	55.0 GB	Q8_0	Comfortable
119 tok/s 71–190 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	55.0 GB	Q8_0	Comfortable
119 tok/s 71–190 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	55.0 GB	Q8_0	Comfortable
119 tok/s 71–190 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	55.0 GB	Q8_0	Comfortable
119 tok/s 71–190 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	55.0 GB	Q8_0	Comfortable
97.9 tok/s 59–157 · low confidence	Tesla PG500-216 NVIDIA	32 GB	1,130 GB/s	Nov 2019	28.0 GB	IQ4_XS	Tight

Speeds are estimates for a single request — one conversation at a time — calculated from memory bandwidth, model size and quantisation. Real throughput varies with the inference runtime and its version. Figures published by hardware vendors measure many simultaneous requests and are much higher.

On record

Full specification

Everything on record for this model. Most of it describes how it was trained rather than how it runs — useful context for judging how much work went into it, and how it compares with models built at a different scale.

Origin

Who built this model, where, and when it was published.

Organisation: AI21 Labs
Organisation type: Industry
Country: Israel
Published: 28 March 2024
Authors: Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avashalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham

What it does

The problem areas the model was built for. A model can carry several of each.

Domain: Language
Task: Language modeling/generation, Chat

Size

How large the model is and how much data it was trained on. Parameters are the figure that decides whether it fits on a given graphics card.

Parameters: 51.6B
Training data: tokens

The training run

What it physically took to train: which chips, how many, for how long, and what that drew from the wall.

Training hardware: NVIDIA H100 SXM5 80GB

Availability

Whether you can obtain the model and run it on your own hardware, which is what decides if any of the graphics-card figures on this page apply.

Weights: Open — downloadable
Model access: Open weights (unrestricted)
Training code: Unreleased
Hugging Face: ai21labs

How it is classified

Labels the source dataset applies when tracking notable models, and how confident it is in the entry.

Record confidence: Confident

Sources

Where this record came from and when it was last checked.

Reference: Jamba: A Hybrid Transformer-Mamba Language Model
Last updated: 28 November 2025

The extremes

The ten fastest GPUs for Jamba

Ranked by estimated tokens per second, newest card first where speeds tie. Because generation is bound by memory bandwidth, this ordering follows bandwidth rather than any gaming benchmark.

The smallest GPUs that still run Jamba

The cheapest route in, by memory capacity. A tight fit runs the model but leaves nothing spare for a longer conversation.

What the numbers mean

Hardware requirements in practice

Minimum card

Radeon PRO V710

Memory needed

24.9 GB

Fastest

282 tok/s

With 51.6B parameters, Jamba lands in the range a serious desktop card can handle once the weights are compressed. 93 of the cards we track can run it.

The entry point is the Radeon PRO V710: 28 GB of memory, Q3_K_M compression, roughly 37.4 tokens per second.

A B200 is the fastest we calculate for it: about 282 tokens per second, from 8,000 GB/s of memory bandwidth.

What this model is

Jamba was published by AI21 Labs, in Israel, in March 2024. It comes out of industry.

It works in Language, and is recorded as doing language modeling/generation, Chat.

The weights are published, so it can be downloaded and run on your own hardware indefinitely, offline, with no account attached. It is published under the ai21labs organisation on Hugging Face.

What decides the speed

Across every card that can run it, the middle of the range is about 71.3 tokens per second, and 89 of them clear the ten tokens per second that roughly matches reading speed.

This is a mixture-of-experts model, which routes each token through only part of itself. It therefore generates far faster than its total size suggests — while still needing every parameter resident in memory, so it is quick without being cheap to hold.

Its internal architecture is not on file, so memory is approximated from the parameter count and marked accordingly. Expect the real figure to differ, more so at long context.

Step by step

How to choose a GPU for Jamba

The table above has already assessed every card we hold specifications for against this model. Getting to your answer takes six steps.

01

Read the memory figure first

The table lists every card that can hold Jamba — around 24.9 GB at Q3_K_M. That figure, not the card's headline performance, is what decides whether it runs.
02

Decide how long your conversations run

Set the context to what you will actually use. The cache grows with the conversation, and it is the usual reason Jamba stops fitting a card that seemed fine.
03

Decide how much compression you will accept

Compression is what makes Jamba fit smaller cards, at some cost in accuracy — Q3_K_M on the smallest card that fits. A minimum quality removes the ones that go too far.
04

Rank by throughput rather than spec sheet

Sort by speed to see how cards rank for Jamba. It will not match a gaming ordering — generation is bound by memory bandwidth, which is why the B200 tops it at 282 tok/s.
05

Check the fit verdict before buying

A tight fit runs Jamba but leaves nothing spare for a longer conversation; comfortable has headroom. If you expect to grow the context, buy for comfortable.
06

See what else that card runs

Each card page repeats this sweep for every model we hold. It answers what else the hardware is good for, beyond Jamba.

Answers

Jamba — common questions

When was Jamba released?

Jamba was published in March 2024. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

What is Jamba used for?

Jamba works in Language, and is recorded as handling language modeling/generation, Chat. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

Where can I download Jamba?

Its weights are published under the ai21labs organisation on Hugging Face. We do not host model files — this site calculates what hardware is needed to run them.

Can I run Jamba if it does not fit in my GPU?

It can be split between the card and system memory, but Jamba generates painfully slowly that way — the nearest miss we calculate is short by 9.4 GB. Nothing on this page assumes offloading.

Would two GPUs run Jamba faster?

Capacity adds across cards; throughput does not. Since 93 of the cards we track already hold Jamba on their own, a second card is rarely the answer here.

Why does the quantisation differ between cards for Jamba?

Because capacity varies, so does how hard Jamba has to be squeezed — 5 distinct levels appear in the table above. Set a minimum quality to compare at one.

How accurate are these Jamba speed estimates?

They are calculated from specifications rather than measured, and each carries a range — 169–452 tok/s on the B200, for instance. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

What GPU do I need to run Jamba?

The smallest card in our catalogue that holds Jamba is the Radeon PRO V710, with 28 GB of memory. It runs the model at Q3_K_M using about 24.9 GB, and produces roughly 37.4 tokens per second. 93 cards in total can run it.

How fast is Jamba on a GPU?

It depends on the card. The quickest we calculate is a B200 at about 282 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and 89 of the cards that can run Jamba clear that.

How much VRAM does Jamba need?

About 24.9 GB at Q3_K_M compression, which is what the smallest card that runs it uses. Less compression needs more: the figures in the memory column above are recalculated for each card, because each one holds the least-compressed version it can.

Is Jamba open source?

Its weights are published, so Jamba can be downloaded and run on your own hardware. Note that open weights is not the same as open source in the full sense — it says nothing about the training data, the training code, or the commercial terms attached.

How many parameters does Jamba have?

Jamba has 51.6B parameters. 51.6B from https://huggingface.co/ai21labs/Jamba-v0.1 (12B active parameters, 52B total available parameters). That figure is the total, and it is what decides how much memory the model needs — roughly half a gigabyte per billion at the compression most people use.

Who created Jamba?

Jamba was published by AI21 Labs, based in Israel, categorised as industry.

Source

Original publication

Record last updated 28 November 2025

The other direction

Looking at it from the other side?

This page starts from the model. If you already own a card and want to know everything it will run, start from the hardware instead.

All AI models