Llemma 34B TPS calculator

Open weights Princeton University,University of Toronto,Vector Institute,University of Cambridge,Carnegie Mellon University (CMU),University of Washington,EleutherAI 34B parameters October 2023

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

132 cards that can run it

818 cards we hold specifications for

Smallest card that fits

RTX A4500

20 GB · Q3_K_M · 21.5 tok/s

Fastest card

B200

99.7 tok/s · 180 GB

Which GPUs can run Llemma 34B?

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

132 cards match

Calculating

					Needs	Quantisation	Fit
99.7 tok/s 60–159 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	37.1 GB	Q8_0	Comfortable
99.7 tok/s 60–159 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	37.1 GB	Q8_0	Comfortable
79.6 tok/s 48–127 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	37.1 GB	Q8_0	Comfortable
79.6 tok/s 48–127 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	37.1 GB	Q8_0	Comfortable
63.6 tok/s 38–102 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	37.1 GB	Q8_0	Comfortable
60.9 tok/s 37–97 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	37.1 GB	Q8_0	Comfortable
60.9 tok/s 37–97 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	37.1 GB	Q8_0	Comfortable
58.3 tok/s 35–93 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	37.1 GB	Q8_0	Comfortable
51.7 tok/s 31–83 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	37.1 GB	Q8_0	Comfortable
51.7 tok/s 31–83 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	37.1 GB	Q8_0	Comfortable
51.7 tok/s 31–83 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	37.1 GB	Q8_0	Comfortable
49.1 tok/s 29–79 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	37.1 GB	Q8_0	Comfortable
41.9 tok/s 25–67 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	37.1 GB	Q8_0	Comfortable
41.9 tok/s 25–67 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	37.1 GB	Q8_0	Comfortable
41.9 tok/s 25–67 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	37.1 GB	Q8_0	Comfortable
41.9 tok/s 25–67 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	37.1 GB	Q8_0	Comfortable
41.9 tok/s 25–67 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	37.1 GB	Q8_0	Comfortable
41.6 tok/s 25–67 · low confidence	DRIVE A100 PROD NVIDIA	32 GB	1,870 GB/s	May 2020	25.2 GB	Q5_K_M	Tight
41.6 tok/s 25–67 · low confidence	GRID A100A NVIDIA	32 GB	1,870 GB/s	May 2020	25.2 GB	Q5_K_M	Tight
39.8 tok/s 24–64 · low confidence	GeForce RTX 5090 NVIDIA	32 GB	1,790 GB/s	Jan 2025	25.2 GB	Q5_K_M	Tight
39.8 tok/s 24–64 · low confidence	GeForce RTX 5090 D NVIDIA	32 GB	1,790 GB/s	Jan 2025	25.2 GB	Q5_K_M	Tight
38.5 tok/s 23–62 · low confidence	GeForce RTX 5090 D V2 NVIDIA	24 GB	1,340 GB/s	Aug 2025	21.3 GB	Q4_K_M	Tight
35.1 tok/s 21–56 · low confidence	A30X NVIDIA	24 GB	1,220 GB/s	Apr 2021	21.3 GB	Q4_K_M	Tight
31.9 tok/s 19–51 · low confidence	Radeon Instinct MI250 AMD	128 GB	3,280 GB/s	Nov 2021	37.1 GB	Q8_0	Comfortable
31.9 tok/s 19–51 · low confidence	Radeon Instinct MI250X AMD	128 GB	3,280 GB/s	Nov 2021	37.1 GB	Q8_0	Comfortable

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: Princeton University,University of Toronto,Vector Institute,University of Cambridge,Carnegie Mellon University (CMU),University of Washington,EleutherAI
Organisation type: Academia,Academia,Academia,Academia,Academia,Academia,Research collective
Country: United States of America, Canada, United Kingdom of Great Britain and Northern Ireland
Published: 16 October 2023
Authors: Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, Sean Welleck

What it does

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

Domain: Mathematics, Language
Task: Mathematical reasoning, Language modeling/generation, Question answering, Code generation
Base model: Code Llama-34B
Numerical format: BF16

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: 34B
Training data: 55,000,000,000 tokens
Epochs: 0.9
Batch size: 4,000,000

Training compute

The arithmetic performed to train the model, measured in floating-point operations. It is a measure of what the training run cost, not of how fast the finished model answers you.

Training compute: 5.4 × 10²³ FLOP
How it was established: Operation counting,Hardware
Fine-tuning compute: 1.3 × 10²² FLOP

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 A100 SXM4 40 GB
Chips used: 256
Chip-hours: 47,000
Power draw: 203.3 kW

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 (restricted use)
Training code: Unreleased
Hugging Face: EleutherAI

How it is classified

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

Record confidence: Confident
Citations: 440

Sources

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

Reference: Llemma: An Open Language Model For Mathematics
Last updated: 25 May 2026

The extremes

The ten fastest GPUs that run Llemma 34B

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 Llemma 34B

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

What you need to run it

Minimum card

RTX A4500

Memory needed

17.3 GB

Fastest

99.7 tok/s

Llemma 34B reaches a parameter count of 34B. That lands in the range a serious desktop card can handle once the weights are compressed. The number of cards we track that can run it: 132.

The least hardware that works is RTX A4500, with a memory capacity of 20 GB, running it at a compression of Q3_K_M and producing around 21.5 tokens per second.

Top of the range is B200, generating roughly 99.7 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.

About this model

Llemma 34B was published by Princeton University,University of Toronto,Vector Institute,University of Cambridge,Carnegie Mellon University (CMU),University of Washington,EleutherAI, in the country recorded as United States of America, during October 2023. The category the publisher falls under is academia,Academia,Academia,Academia,Academia,Academia,Research collective.

It works in the domain of Mathematics, Language, and is recorded as performing the task of mathematical reasoning, Language modeling/generation, Question answering, Code generation.

Its starting point was an existing base model, Code Llama-34B. That is why it shares the base model's general shape and size.

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

How fast it runs, and why

Across every card that can run it, the middle of the range sits at 21.2 tokens per second. Producing text faster than most people read it: 103 of them.

Being dense, it reads all of itself per token, which is why the ordering by speed below follows the ordering by memory bandwidth so closely.

Without the attention layout on record, the memory column is an approximation. It is close enough to choose hardware by, and least reliable at long context.

How it was trained

Training it took a computation budget of roughly 5.4 × 10²³ FLOP, on hardware recorded as NVIDIA A100 SXM4 40 GB. That figure measures what producing the model cost, and has no bearing on how fast it answers.

It was trained on a corpus of about 55,000,000,000 tokens of text.

Step by step

How to choose a GPU for Llemma 34B

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

Start from what it actually needs, which is the requirement of Llemma 34B, needing around 17.3 GB at a compression of Q3_K_M. No amount of processing power compensates for a card that cannot hold it.
02

Set the context length you will work at

Longer conversations cost memory on top of what the weights need. Move the slider to your real working length before trusting any row for Llemma 34B.
03

Set a quality floor

Each card runs the least-compressed copy it can hold, reaching a compression of Q3_K_M on the smallest card that fits. Setting a minimum quality drops the cards that only manage it by squeezing further than you would want, and holds the comparison at one level.
04

Compare tokens per second, not specifications

Sort by speed to see how cards rank for Llemma 34B. It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B200, at 99.7 tok/s.
05

Check the fit verdict before buying

Tight means it loads and works with no room to raise the context later, in the case of Llemma 34B. Comfortable means you can grow the context later. That difference matters more than a few tokens per second, so buy for comfortable if you expect to.
06

Open the card you have settled on

Every card name links to its own page, which runs the same calculation across the whole model catalogue. A card is usually bought for more than one model, so it is worth a look before buying for Llemma 34B.

Answers

Llemma 34B — common questions

Llemma 34B— can I run it on a GPU holding 24 GB?

Yes. The card GeForce RTX 5090 D V2, holding 24 GB, runs it at a compression of Q4_K_M, using about 21.3 GB and generating roughly 38.5 tokens per second. The fit is tight.

Llemma 34B— is it open source?

Its weights are published, so it 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.

Llemma 34B— how many parameters does it have?

It has a parameter count of 34B. https://arxiv.org/abs/2310.10631. 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.

Llemma 34B— who created it?

It was published by Princeton University,University of Toronto,Vector Institute,University of Cambridge,Carnegie Mellon University (CMU),University of Washington,EleutherAI, based in United States of America, an organisation categorised as academia,Academia,Academia,Academia,Academia,Academia,Research collective.

Llemma 34B— when was it released?

It was published in October 2023. Capability per parameter has improved considerably since, so a newer model of the same size is often the better use of the same hardware.

Llemma 34B— what is it used for?

It works in the domain of Mathematics, Language, and is recorded as handling the task of mathematical reasoning, Language modeling/generation, Question answering, Code generation. Models frequently carry more than one of each, and the tags describe purpose rather than capability limits.

Llemma 34B— where can I download it?

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

Llemma 34B— how much compute was used to train it?

Training consumed around 5.4 × 10²³ FLOP, on hardware recorded as NVIDIA A100 SXM4 40 GB. That measures what producing the model cost and says nothing about how quickly it answers once trained — inference speed comes from memory bandwidth, not from the training budget.

Llemma 34B— can I run it if it does not fit in my GPU?

Only by offloading, which is usually a false economy: the part held in system memory drags the whole thing down. The nearest miss we calculate falls short by 6.9 GB. Every figure here assumes the whole model is resident on the card.

Llemma 34B— would two GPUs run it faster?

Two cards buy memory rather than speed, which matters only if one card cannot hold it. The number that can: 132. So a second card is rarely the answer here.

Llemma 34B— why does the quantisation differ between cards?

Each card is shown running the least-compressed copy it can hold. The number of distinct compression levels across the cards that fit it: 5. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.

Llemma 34B— how accurate are these speed estimates?

These are estimates with real error bars, and any of them could reasonably land anywhere in its published range depending on which runtime you use. The fastest result here: 60–159 tok/s on B200. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

Llemma 34B— what GPU do I need to run it?

The smallest card in our catalogue that holds it is RTX A4500, with a memory capacity of 20 GB. It runs the model at a compression of Q3_K_M using about 17.3 GB, and produces roughly 21.5 tokens per second. The number of cards able to run it in total: 132.

Llemma 34B— how fast is it on a GPU?

It depends on the card. The quickest we calculate is B200, at about 99.7 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and the number of cards clearing that: 103.

Llemma 34B— how much VRAM does it need?

It needs about 17.3 GB at a compression of Q3_K_M, 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.

Source

Original publication

Record last updated 25 May 2026

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