Tulu 3 405B TPS calculator

Open weights Allen Institute for AI,University of Washington 405B parameters January 2025

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

4 cards that can run it

818 cards we hold specifications for

Smallest card that fits

Radeon Instinct MI325X

256 GB · IQ4_XS · 12.0 tok/s

Fastest card

B300

19.3 tok/s · 288 GB

Which GPUs can run Tulu 3 405B?

Set the inputs, read the answer

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

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

4 cards match

Calculating
Needs Quantisation Fit
19.3 tok/s

12–31 · low confidence

B300 NVIDIA 288 GB 8,000 GB/s Sep 2025 245.7 GB Q4_K_M Tight
15.4 tok/s

9–25 · low confidence

Radeon Instinct MI350X AMD 288 GB 8,190 GB/s Jan 2025 245.7 GB Q4_K_M Tight
15.4 tok/s

9–25 · low confidence

Radeon Instinct MI355X AMD 288 GB 8,190 GB/s Jan 2025 245.7 GB Q4_K_M Tight
12.0 tok/s

7–19 · low confidence

Radeon Instinct MI325X AMD 256 GB 6,000 GB/s Oct 2024 222.1 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
Allen Institute for AI,University of Washington
Organisation type
Research collective,Academia
Country
United States of America
Published
30 January 2025
Authors
Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Chris Wilhelm, Luca Soldaini, Noah A. Smith, Yizhong Wang, Pradeep Dasigi, Hannaneh Hajishirzi

What it does

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

Domain
Language
Task
Language modeling/generation, Question answering, Quantitative reasoning
Base model
Llama 3.1-405B

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
405B

405B

Training data
tokens

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.

How it was established
Hardware
Fine-tuning compute
3.8 × 10²² FLOP

SFT and DPO infrastructure usage is unknown RLVR: "the final 405B RL run takes 46 hours on 256 GPUs" 989400000000000 FLOP / GPU / sec [H100 reported, bf16 assumed] * 256 GPUs * 46 hours * 3600 sec / hour * 0.3 [assumed utilization] = 1.2583268e+22 FLOP Assuming that SFT and DPO training compute was same OOM as RLVR stage (similar to 70B model), 1.2583268e+22 FLOP * 3 = 3.7749804e+22 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 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 (restricted use)
Training code
Open source

Llama 3.1 license https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B Apache 2.0 https://github.com/allenai/open-instruct

Hugging Face
allenai

How it is classified

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

Record confidence
Likely

Sources

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

Reference
Tulu 3: Pushing Frontiers in Open Language Model Post-Training
Last updated
28 November 2025

The extremes

The ten fastest GPUs that run Tulu 3 405B

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.

  1. 01 B300 288 GB · 8,000 GB/s · Q4_K_M 19.3 tok/s
  2. 02 Radeon Instinct MI350X 288 GB · 8,190 GB/s · Q4_K_M 15.4 tok/s
  3. 03 Radeon Instinct MI355X 288 GB · 8,190 GB/s · Q4_K_M 15.4 tok/s
  4. 04 Radeon Instinct MI325X 256 GB · 6,000 GB/s · IQ4_XS 12.0 tok/s

What the numbers mean

The hardware side

Minimum card

Radeon Instinct MI325X

Memory needed

222.1 GB

Fastest

19.3 tok/s

Tulu 3 405B reaches a parameter count of 405B. That is beyond what any single graphics card holds. Running it means either splitting it across several cards or renting hardware built for the job, and every card able to hold it alone is a datacentre part. The number that can: 4.

At the low end it is handled by Radeon Instinct MI325X, with a memory capacity of 256 GB, running it at a compression of IQ4_XS and producing around 12.0 tokens per second.

At the other end sits B300, generating roughly 19.3 tokens per second on the strength of a memory bandwidth of 8,000 GB/s.

What this model is

Tulu 3 405B was published by Allen Institute for AI,University of Washington, in the country recorded as United States of America, during January 2025. It comes out of an organisation categorised as research collective,Academia.

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

Rather than being trained from scratch, it is derived from Llama 3.1-405B. Most models at this scale are adapted from an existing base rather than built from nothing.

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 allenai.

What decides the speed

Half the cards that hold it manage more than 15.4 tokens per second. Exceeding reading speed outright: 4 of them.

Every weight participates in every token here, so bandwidth is the whole story: the ranking below is effectively a ranking of memory throughput.

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 Tulu 3 405B

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

  1. 01

    Start from the memory column

    Start from what it actually needs, which is the requirement of Tulu 3 405B, needing around 222.1 GB at a compression of IQ4_XS. No amount of processing power compensates for a card that cannot hold it.

  2. 02

    Set the context length you will work at

    The conversation occupies memory too, and grows as it goes. Set the slider to the length you expect, because at long context a card that handles short questions easily can be dropped by Tulu 3 405B.

  3. 03

    Choose how far you will compress it

    Compression is what makes a model fit smaller cards, at some cost in accuracy, reaching a compression of IQ4_XS 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.

  4. 04

    Rank by throughput rather than spec sheet

    The speed ordering is effectively an ordering by memory bandwidth, for Tulu 3 405B. It will not match a gaming ordering, because generation is bound by memory bandwidth. The card topping the list is B300, at 19.3 tok/s.

  5. 05

    Look at the headroom, not just the fit

    A tight fit runs, but leaves nothing spare for a longer conversation, in the case of Tulu 3 405B. 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.

  6. 06

    See what else that card runs

    Following a card through to its own page shows every other model it can hold, which is the question that follows once you have settled on Tulu 3 405B.

Answers

Tulu 3 405B — common questions

01

Tulu 3 405B— what is it used for?

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

02

Tulu 3 405B— where can I download it?

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

03

Tulu 3 405B— 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 72.9 GB. Every figure here assumes the whole model is resident on the card.

04

Tulu 3 405B— 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: 4. So a second card is rarely the answer here.

05

Tulu 3 405B— why does the quantisation differ between cards?

Because capacity varies, so does how hard it has to be squeezed. The number of distinct levels in the table above: 2. Bigger cards get the more accurate version, and the quality floor above pins the comparison to one level.

06

Tulu 3 405B— 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: 12–31 tok/s on B300. The same model and card vary by thirty to fifty per cent depending on the inference software and its version.

07

Tulu 3 405B— what GPU do I need to run it?

The smallest card in our catalogue that holds it is Radeon Instinct MI325X, with a memory capacity of 256 GB. It runs the model at a compression of IQ4_XS using about 222.1 GB, and produces roughly 12.0 tokens per second. The number of cards able to run it in total: 4.

08

Tulu 3 405B— how fast is it on a GPU?

It depends on the card. The quickest we calculate is B300, at about 19.3 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: 4.

09

Tulu 3 405B— how much VRAM does it need?

It needs about 222.1 GB at a compression of IQ4_XS, 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.

10

Tulu 3 405B— 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.

11

Tulu 3 405B— how many parameters does it have?

It has a parameter count of 405B. 405B. 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.

12

Tulu 3 405B— who created it?

It was published by Allen Institute for AI,University of Washington, based in United States of America, an organisation categorised as research collective,Academia.

13

Tulu 3 405B— when was it released?

It was published in January 2025.

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.