SAVRN
Search Contact SAVRN

SAVRN Model Hub · Comparisons

blip2-opt-2.7b vs Qwen2.5-VL-3B-Instruct

Blip2-opt-2.7b has 3.7B parameters and Qwen2.5-VL-3B-Instruct has 3.8B parameters; at 16-bit, blip2-opt-2.7b needs about 9 GB (1x MI300X from $1.85 an hour) and Qwen2.5-VL-3B-Instruct about 9 GB (1x MI300X from $1.85 an hour).

Published metadata for 2 models, each read from its own repository.
Field blip2-opt-2.7b
Salesforce/blip2-opt-2.7b
Qwen2.5-VL-3B-Instruct
Qwen/Qwen2.5-VL-3B-Instruct
Publisher Salesforce AI Research Qwen
Task Image and text to text Image and text to text
Modality Image and text Image and text
Parameters, as reported 3.7B parameters 3.8B parameters
Architecture Blip2ForConditionalGeneration Qwen2_5_VLForConditionalGeneration
Library transformers transformers
Context length Not stated 128,000 tokens
Repository size 30.5 GB 7.5 GB
Artifact formats safetensors, pytorch safetensors
License mit Not stated
Access Open weights, no gate Open weights, no gate
Memory at 16-bit (weights and margin) 9 GB 9 GB
Cheapest GPUs at 16-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Memory at 4-bit (weights and margin) 2.2 GB 2.3 GB
Cheapest GPUs at 4-bit, per hour 1x MI300X, $1.85 1x MI300X, $1.85
Revision viewed 59a1ef6c1e51 66285546d2b8
Downloads reported by the hub 662.5k 2.5M
Last observed 2026-09-18 2026-09-18

An evaluation row appears only where at least two of these models report the same benchmark with the same stated configuration, metric, unit and setup. Different evaluators stay named in each cell. Values are shown as reported: no unit conversion, no ranking.

Other Reported Results

These results are listed for each model on its own, because the conditions needed to compare them are not stated or do not match. Two results that leave a condition blank are not assumed to share it.

Qwen2.5-VL-3B-Instruct

BenchmarkConditionsResultReported byRevisionDate
MMMU/MMMU_Pro Task mmmu_pro_visionMetric mmmu_pro_visionComparison conditions not established 32.7 Model Card
Reported by a third party
Evaluated revision not stated 2026-01-28
likaixin/ScreenSpot-Pro Task android_studio_macosMetric android_studio_macosComparison conditions not established 2.5 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task autocad_windowsMetric autocad_windowsComparison conditions not established 2.9 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task blender_windowsMetric blender_windowsComparison conditions not established 21.1 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task davinci_macosMetric davinci_macosComparison conditions not established 11.4 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task eviews_windowsMetric eviews_windowsComparison conditions not established 62 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task excel_macosMetric excel_macosComparison conditions not established 14.1 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task fruitloops_windowsMetric fruitloops_windowsComparison conditions not established 15.8 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task illustrator_windowsMetric illustrator_windowsComparison conditions not established 0 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task inventor_windowsMetric inventor_windowsComparison conditions not established 1.4 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task linux_common_linuxMetric linux_common_linuxComparison conditions not established 4 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task macos_common_macosMetric macos_common_macosComparison conditions not established 9.2 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task matlab_macosMetric matlab_macosComparison conditions not established 23.7 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task origin_windowsMetric origin_windowsComparison conditions not established 8.1 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task overallMetric overallComparison conditions not established 16.1 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task photoshop_windowsMetric photoshop_windowsComparison conditions not established 23.5 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task powerpoint_windowsMetric powerpoint_windowsComparison conditions not established 19.5 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task premiere_windowsMetric premiere_windowsComparison conditions not established 11.5 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task pycharm_macosMetric pycharm_macosComparison conditions not established 10.3 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task quartus_windowsMetric quartus_windowsComparison conditions not established 15.6 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task solidworks_windowsMetric solidworks_windowsComparison conditions not established 3.9 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task stata_windowsMetric stata_windowsComparison conditions not established 10.2 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task unreal_engine_windowsMetric unreal_engine_windowsComparison conditions not established 25.7 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task vivado_windowsMetric vivado_windowsComparison conditions not established 17.5 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task vmware_macosMetric vmware_macosComparison conditions not established 19.5 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task vscode_macosMetric vscode_macosComparison conditions not established 20 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task windows_common_windowsMetric windows_common_windowsComparison conditions not established 4.9 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17
likaixin/ScreenSpot-Pro Task word_macosMetric word_macosComparison conditions not established 51.2 ScreenSpot-Pro Leaderboard
Reported by a third party
Evaluated revision not stated 2026-03-17

SAVRN's Notes on blip2-opt-2.7b

Nine gigabytes of memory is the number to plan around, which is what the 16-bit run needs with 7.5 GB of weights loaded. That puts captioning and image question answering on one 192 GB MI300X at $1.85 per hour on demand, the cheapest setup we track, with most of the card still free. Mind the gap between that figure and the download: the files total 30.5 GB because the weights are stored in float32, so size storage to the larger number.

MIT is about as light as a license gets. Commercial use, modification and redistribution are permitted, and the only obligation is keeping the copyright and permission notices with the files. Two checks: the release date is 2023-02-06 and the last update 2025-02-03, so weigh it against newer image-and-text models, and our file lists no context length, so test your own prompt and image sizes.

SAVRN's Notes on Qwen2.5-VL-3B-Instruct

Screenshots, scanned drawings and video frames go in with a prompt, text comes out, and the 128,000-token context is where the pictures live. At 3.8 billion parameters it runs in 9.0 GB at 16-bit and 2.3 GB at 4-bit, so the 192 GB MI300X we list at $1.85 an hour is more card than the weights need. Give it that card anyway for the context: filling 128,000 tokens with frames costs memory the weights figure hides, and the configuration carries a 32,768-token sliding window your serving stack must honor.

Read the license first. Our facts hold no summary of its terms; the publisher labels it qwen-research and links its own license file, so nobody should assume commercial use before reading the full text. Then test it on your own screens: ScreenSpot-Pro reports per-application scores from 2.5 to 62.0, third-party reported, too wide a spread to buy hardware against.

Questions

Which is larger, blip2-opt-2.7b or Qwen2.5-VL-3B-Instruct?

Qwen2.5-VL-3B-Instruct (3.8B parameters) is larger than blip2-opt-2.7b (3.7B parameters), by the parameter counts their publishers report.

Which is cheaper to run, blip2-opt-2.7b or Qwen2.5-VL-3B-Instruct?

At 4-bit, blip2-opt-2.7b fits on 1x MI300X from $1.85 an hour and Qwen2.5-VL-3B-Instruct on 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use blip2-opt-2.7b commercially?

Yes. blip2-opt-2.7b is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

Related Comparisons