BLIP-2 (Q-Former) TPS calculator

Open weights Salesforce Research 1.5B parameters January 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

818 cards that can run it

818 cards we hold specifications for

Smallest card that fits

Tesla C1080

4 GB · Q8_0 · 24.9 tok/s

Fastest card

B200

2,289 tok/s · 180 GB

Which GPUs can run BLIP-2 (Q-Former)?

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

818 cards match

Calculating

					Needs	Quantisation	Fit
2,289 tok/s 1,374–3,663 · low confidence	B200 NVIDIA	180 GB	8,000 GB/s	Jan 2024	2.3 GB	Q8_0	Comfortable
2,289 tok/s 1,374–3,663 · low confidence	B300 NVIDIA	288 GB	8,000 GB/s	Sep 2025	2.3 GB	Q8_0	Comfortable
1,828 tok/s 1,097–2,925 · low confidence	Radeon Instinct MI350X AMD	288 GB	8,190 GB/s	Jan 2025	2.3 GB	Q8_0	Comfortable
1,828 tok/s 1,097–2,925 · low confidence	Radeon Instinct MI355X AMD	288 GB	8,190 GB/s	Jan 2025	2.3 GB	Q8_0	Comfortable
1,462 tok/s 877–2,339 · low confidence	Radeon Instinct MI300 AMD	128 GB	6,550 GB/s	Jan 2023	2.3 GB	Q8_0	Comfortable
1,399 tok/s 840–2,239 · low confidence	H200 NVL NVIDIA	141 GB	4,890 GB/s	Nov 2024	2.3 GB	Q8_0	Comfortable
1,399 tok/s 840–2,239 · low confidence	H200 SXM 141 GB NVIDIA	141 GB	4,890 GB/s	Nov 2024	2.3 GB	Q8_0	Comfortable
1,339 tok/s 804–2,143 · low confidence	Radeon Instinct MI325X AMD	256 GB	6,000 GB/s	Oct 2024	2.3 GB	Q8_0	Comfortable
1,189 tok/s 713–1,902 · low confidence	Radeon Instinct MI300A AMD	128 GB	5,325 GB/s	Dec 2023	2.3 GB	Q8_0	Comfortable
1,189 tok/s 713–1,902 · low confidence	Radeon Instinct MI300X AMD	192 GB	5,325 GB/s	Dec 2023	2.3 GB	Q8_0	Comfortable
1,189 tok/s 713–1,902 · low confidence	Radeon Instinct MI308X AMD	192 GB	5,325 GB/s	Dec 2023	2.3 GB	Q8_0	Comfortable
1,128 tok/s 677–1,804 · low confidence	H100 NVL 94 GB NVIDIA	94 GB	3,940 GB/s	Mar 2023	2.3 GB	Q8_0	Comfortable
962 tok/s 577–1,538 · low confidence	H100 PCIe 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	2.3 GB	Q8_0	Comfortable
962 tok/s 577–1,538 · low confidence	H100 SXM5 80 GB NVIDIA	80 GB	3,360 GB/s	Oct 2022	2.3 GB	Q8_0	Comfortable
962 tok/s 577–1,538 · low confidence	H100 SXM5 94 GB NVIDIA	94 GB	3,360 GB/s	Mar 2023	2.3 GB	Q8_0	Comfortable
962 tok/s 577–1,538 · low confidence	H100 SXM5 96 GB NVIDIA	96 GB	3,360 GB/s	Mar 2023	2.3 GB	Q8_0	Comfortable
962 tok/s 577–1,538 · low confidence	H800 SXM5 NVIDIA	80 GB	3,360 GB/s	Mar 2023	2.3 GB	Q8_0	Comfortable
732 tok/s 439–1,171 · low confidence	Radeon Instinct MI250 AMD	128 GB	3,280 GB/s	Nov 2021	2.3 GB	Q8_0	Comfortable
732 tok/s 439–1,171 · low confidence	Radeon Instinct MI250X AMD	128 GB	3,280 GB/s	Nov 2021	2.3 GB	Q8_0	Comfortable
610 tok/s 366–976 · low confidence	Data Center GPU Max 1550 Intel	128 GB	3,280 GB/s	Jan 2023	2.3 GB	Q8_0	Comfortable
597 tok/s 358–955 · low confidence	Data Center GPU Max Subsystem Intel	128 GB	3,210 GB/s	Jan 2023	2.3 GB	Q8_0	Comfortable
584 tok/s 350–934 · low confidence	A100 SXM4 80 GB NVIDIA	80 GB	2,040 GB/s	Nov 2020	2.3 GB	Q8_0	Comfortable
584 tok/s 350–934 · low confidence	A100X NVIDIA	80 GB	2,040 GB/s	Jun 2021	2.3 GB	Q8_0	Comfortable
584 tok/s 350–934 · low confidence	A800 SXM4 80 GB NVIDIA	80 GB	2,040 GB/s	Aug 2022	2.3 GB	Q8_0	Comfortable
584 tok/s 350–934 · low confidence	H100 CNX NVIDIA	80 GB	2,040 GB/s	Mar 2023	2.3 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: Salesforce Research
Organisation type: Industry
Country: United States of America
Published: 30 January 2023
Authors: Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi

What it does

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

Domain: Vision, Language
Task: Visual question answering, Image captioning
Numerical format: FP16

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: 1.5B
Training data: 2,322,000,000 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.

Training compute: 1.2 × 10²¹ FLOP
How it was established: Hardware

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: 16
Chip-hours: 3,200
Wall-clock time: 200 hours (8.3 days)
Power draw: 12.8 kW
Compute cost: $1,961

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: Open source

How it is classified

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

Why it is tracked: SOTA improvement
Record confidence: Confident
Citations: 8,019

Sources

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

Reference: BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
Last updated: 25 May 2026

The extremes

The ten fastest GPUs that run BLIP-2 (Q-Former)

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 BLIP-2 (Q-Former)

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

Tesla C1080

Memory needed

2.3 GB

Fastest

2,289 tok/s

BLIP-2 (Q-Former) is small enough at 1.5B parameters that hardware is rarely the obstacle — 818 of the cards we track can run it, including cards several years old.

The entry point is the Tesla C1080: 4 GB of memory, Q8_0 compression, roughly 24.9 tokens per second.

The quickest result comes from a B200 at around 2,289 tokens per second — its 8,000 GB/s of bandwidth is what buys that.

What this model is

BLIP-2 (Q-Former) was published by Salesforce Research, in United States of America, in January 2023. industry is the category the publisher falls under.

It works in Vision, Language, and is recorded as doing visual question answering, Image captioning.

The weights being open is what puts this page in the calculator rather than only in the catalogue: it is a model you can actually hold.

What decides the speed

The median result is around 64.3 tokens per second; 796 cards produce text faster than most people read it.

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

Memory here is estimated from size rather than computed from the architecture, which is not recorded for this model — the numbers are indicative rather than exact.

Training and provenance

Training it took roughly 1.2 × 10²¹ FLOP of computation, on NVIDIA A100 SXM4 40 GB — a measure of what producing the model cost, not of how fast it answers.

It was trained on about 2,322,000,000 tokens of text.

Its inclusion criterion is sOTA improvement.

Step by step

How to choose a GPU for BLIP-2 (Q-Former)

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

01

Start from the memory column

Every card here has been checked against BLIP-2 (Q-Former) — around 2.3 GB at Q8_0. Capacity is the gate — a card either holds it or it does not.
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 BLIP-2 (Q-Former) stops fitting a card that seemed fine.
03

Choose how far you will compress it

Each card runs the least-compressed copy it can hold — Q8_0 on the smallest card that fits. Setting a floor drops the cards that only manage BLIP-2 (Q-Former) by squeezing it further than you would want.
04

Compare tokens per second, not specifications

Sort by speed to see how cards rank for BLIP-2 (Q-Former). It will not match a gaming ordering — generation is bound by memory bandwidth, which is why the B200 tops it at 2,289 tok/s.
05

Check the fit verdict before buying

Tight means BLIP-2 (Q-Former) loads and works, with no room to raise the context later. Comfortable means you can. The difference matters more than a few tokens per second.
06

Check the card from the other side

Following a card through to its own page shows every other model it can hold, which is the question that follows once BLIP-2 (Q-Former) is settled.

Answers

BLIP-2 (Q-Former) — common questions

When was BLIP-2 (Q-Former) released?

BLIP-2 (Q-Former) was published in January 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.

What is BLIP-2 (Q-Former) used for?

BLIP-2 (Q-Former) works in Vision, Language, and is recorded as handling visual question answering, Image captioning. A model can carry several of each, so these are the areas it was built for rather than a limit on what it will attempt.

Where can I download BLIP-2 (Q-Former)?

The weights for BLIP-2 (Q-Former) are published, though we do not hold a repository link for it. This site calculates hardware requirements rather than hosting model files.

How much compute was used to train BLIP-2 (Q-Former)?

Around 1.2 × 10²¹ FLOP, on 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.

Can I run BLIP-2 (Q-Former) if it does not fit in my GPU?

Only by offloading, which is usually a false economy: the part in system memory drags the whole thing down. Our figures for BLIP-2 (Q-Former) assume it is fully resident.

Would two GPUs run BLIP-2 (Q-Former) faster?

A second card roughly doubles the memory available but not the generation rate. With 818 cards already able to run BLIP-2 (Q-Former) alone, the case for pairing is weak.

Why does the quantisation differ between cards for BLIP-2 (Q-Former)?

Each card is shown running the least-compressed copy it can hold, and BLIP-2 (Q-Former) appears at 1 different compression levels across the cards that fit it. Bigger cards get the more accurate version.

How accurate are these BLIP-2 (Q-Former) speed estimates?

Every figure is derived from memory bandwidth and model size, not benchmarked. That is why each is published as a range such as 1,374–3,663 tok/s on the B200 rather than a single number.

What GPU do I need to run BLIP-2 (Q-Former)?

The smallest card in our catalogue that holds BLIP-2 (Q-Former) is the Tesla C1080, with 4 GB of memory. It runs the model at Q8_0 using about 2.3 GB, and produces roughly 24.9 tokens per second. 818 cards in total can run it.

How fast is BLIP-2 (Q-Former) on a GPU?

It depends on the card. The quickest we calculate is a B200 at about 2,289 tokens per second; the slowest that still runs it manages considerably less. Reading speed is around ten tokens per second, and 796 of the cards that can run BLIP-2 (Q-Former) clear that.

How much VRAM does BLIP-2 (Q-Former) need?

About 2.3 GB at Q8_0 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.

Can I run BLIP-2 (Q-Former) on a 8 GB GPU?

Yes. A CMP 170HX 8 GB with 8 GB runs it at Q8_0, using about 2.3 GB and generating roughly 426 tokens per second — a comfortable fit.

Can I run BLIP-2 (Q-Former) on a 12 GB GPU?

Yes. A GeForce RTX 3080 Ti with 12 GB runs it at Q8_0, using about 2.3 GB and generating roughly 261 tokens per second — a comfortable fit.

Can I run BLIP-2 (Q-Former) on a 16 GB GPU?

Yes. A Tesla V100 SXM2 16 GB with 16 GB runs it at Q8_0, using about 2.3 GB and generating roughly 323 tokens per second — a comfortable fit.

Can I run BLIP-2 (Q-Former) on a 24 GB GPU?

Yes. A GeForce RTX 5090 D V2 with 24 GB runs it at Q8_0, using about 2.3 GB and generating roughly 383 tokens per second — a comfortable fit.

Is BLIP-2 (Q-Former) open source?

Its weights are published, so BLIP-2 (Q-Former) 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 BLIP-2 (Q-Former) have?

BLIP-2 (Q-Former) has 1.5B parameters. Q-Former has 188M params. The BLIP-2 system overall has "54x fewer trainable parameters" than Flamingo80B. 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 BLIP-2 (Q-Former)?

BLIP-2 (Q-Former) was published by Salesforce Research, based in United States of America, categorised as industry.

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