Haytham sees what is there.

A compact vision model that answers questions about images accurately, quickly, and on a single GPU.

See the results

Ask. Look. Answer.

  • VQAv2 pairs

    81.3%

    #1 of 9 models

  • MME

    89.9%

    #1 of 9 models

  • POPE

    88.4%

    Accuracy

  • Parameters

    6.8B

    #1 of 9 models

  • Peak memory

    18.7 GB

    #2 of 9 models

  • Forward latency

    163 ms

    One image, H100

Same prompts and hardware for all 9 models.

Tested on five benchmarks.

Answers follow the image.

VQAv2 pairs check whether the answer changes when the image does.

VQAv2 pairs: pair accuracy by model, highest first
ModelPair accuracy (%)
Haytham v381.3
Clef 27B79.2
Clef Omni78.2
Clef Flash77.2
Qwen3-VL-8B77.0
MiniCPM-V-4.576.9
Qwen3.5-9B76.5
InternVL3.5-8B74.7
Gemma-4-12B69.0
VQAv2 pairs · pair accuracy · 5,000 complementary pairs

Broad perception and reasoning.

MME covers fourteen perception and reasoning skills.

MME: question accuracy by model, highest first
ModelQuestion accuracy (%)
Haytham v389.9
Qwen3-VL-8B89.8
Clef Omni89.6
MiniCPM-V-4.589.2
Clef 27B89.0
InternVL3.5-8B88.1
Qwen3.5-9B87.3
Clef Flash86.6
Gemma-4-12B77.8
MME · question accuracy · 2,374 questions

Grounded in the image.

POPE checks whether the model reports objects that are not in the image.

POPE: accuracy by model, highest first
ModelAccuracy (%)
Qwen3-VL-8B89.2
Clef Omni88.7
Qwen3.5-9B88.7
Haytham v388.4
InternVL3.5-8B88.4
Clef 27B87.9
MiniCPM-V-4.587.2
Clef Flash87.0
Gemma-4-12B84.4
POPE · accuracy · 9,000 questions

Right four times over.

NaturalBench asks four linked questions, and every answer has to be right.

NaturalBench: group accuracy by model, highest first
ModelGroup accuracy (%)
Clef 27B40.7
Haytham v340.0
Clef Flash37.8
Qwen3.5-9B37.0
Clef Omni35.3
Qwen3-VL-8B33.6
MiniCPM-V-4.533.1
Gemma-4-12B29.1
InternVL3.5-8B28.3
NaturalBench · group accuracy · 1,450 groups of four

Charts are the hard part.

HallusionBench uses charts and edited figures that trip up models leaning on language priors.

HallusionBench: question accuracy by model, highest first
ModelQuestion accuracy (%)
Clef 27B79.2
Clef Flash78.4
Gemma-4-12B73.7
Qwen3-VL-8B73.1
Clef Omni72.8
Qwen3.5-9B72.3
MiniCPM-V-4.572.2
Haytham v371.4
InternVL3.5-8B70.5
HallusionBench · question accuracy · 951 image questions

Charts get a specialist.

Haytham routes charts, maps and documents to a specialist model and lifts HallusionBench. Photos still answer in tens of milliseconds.

HallusionBench

51.5 ms per photo · 75.8 ms handed off · 512 image tokens, one H100, median

+5.6 pts from hand-off · 95% CI +3.3 to +7.9

Other four benchmarks: unchanged
Benchmarkv3Hand-offChange
POPE88.488.40.0
VQAv2 pairs81.381.5+0.2
NaturalBench40.040.1+0.1
MME89.989.8−0.1
HallusionBench question accuracy, Haytham with and without hand-off
ModelAccuracy (%)
Haytham v371.4
Haytham with hand-off77.0
HallusionBench · question accuracy · 951 image questions

Small footprint. Fast answers.

Haytham runs on one GPU with modest memory and responds quickly to a single image.

All nine models, memory and latency
ModelGBForward msEnd-to-end ms
InternVL3.5-8B18.1141175
Haytham v318.7163239
MiniCPM-V-4.520.6127225
Qwen3.5-9B21.4218254
Qwen3-VL-8B21.6205256
Clef Flash21.8211289
Gemma-4-12B24.24464
Clef 27B58.3441514
Clef Omni64.0101156
Single image, one H100, batch size 1. Peak allocated GPU memory during timed runs, weights included.

Every model saw the same items on one H100 in bf16 at batch size 1, using our own prompt protocol, so results are not comparable to public leaderboards.

Vision, made practical.

Explore ASIC Intelligence Explore ASIC Research