Tiny vision-language models for visual document retrieval, built on UltraBERT.
Athrael Soju PRO
athrael-soju
AI & ML interests
Yes
Recent Activity
liked a dataset 1 day ago
whybe-choi/ko-vdr-hn updated a dataset 1 day ago
athrael-soju/vdr-train-500k updated a model 1 day ago
athrael-soju/ColUltraVBERTOrganizations
-
athrael-soju/colqwen3.5-4.5B-v3
Visual Document Retrieval • 5B • Updated • 5k • 17 -
athrael-soju/HydraQwen3.5-0.8B
Visual Document Retrieval • Updated • 14 -
athrael-soju/ColGemma4-E4B-IT-Base
Visual Document Retrieval • Updated • 1 -
athrael-soju/ColGemma4-E2B-IT-Base
Visual Document Retrieval • Updated • 1
Visual document retrieval models on Qwen3.5 backbone. ViDoRe v3 leaderboard competitors, 128-dim multi-vector.
-
The Price of Anarchy in Disaggregated Inference
Paper • 2606.17081 • Published • 1 -
Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation
Paper • 2512.02660 • Published • 1 -
Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs: A Comprehensive Benchmark and Deployment Study
Paper • 2603.10031 • Published -
Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model
Paper • 2603.28554 • Published
-
vultr/VultronCoderAtlas-Kimi-K3-2T
Image-Text-to-Text • 2.1T • Updated • 177 • 3 -
vultr/VultronRetrieverPrime-Qwen3.5-8B
Visual Document Retrieval • 8B • Updated • 6.87k • 7 -
vultr/VultronRetrieverCore-Qwen3.5-4.5B
Visual Document Retrieval • 5B • Updated • 266 • 4 -
vultr/VultronRetrieverFlash-Qwen3.5-0.8B
Visual Document Retrieval • 0.9B • Updated • 2.76k • 6
Dual-head VLM: ColBERT retrieval + autoregressive generation by toggling one LoRA. Canonical 4B + 0.8B, omni proof-of-concept, baselines.
-
athrael-soju/HydraQwen3.5-4B
Visual Document Retrieval • Updated • 12 -
athrael-soju/HydraQwen3.5-0.8B
Visual Document Retrieval • Updated • 14 -
athrael-soju/HydraQwen2.5-Omni-3B
Visual Document Retrieval • Updated • 2 -
athrael-soju/DualHead-GritLM-Qwen3.5-4B
Visual Document Retrieval • Updated • 17 • 1
ColBERT-style late-interaction visual document retrieval adapters built on Google Gemma-4 (E2B and E4B variants).
Tiny vision-language models for visual document retrieval, built on UltraBERT.
-
vultr/VultronCoderAtlas-Kimi-K3-2T
Image-Text-to-Text • 2.1T • Updated • 177 • 3 -
vultr/VultronRetrieverPrime-Qwen3.5-8B
Visual Document Retrieval • 8B • Updated • 6.87k • 7 -
vultr/VultronRetrieverCore-Qwen3.5-4.5B
Visual Document Retrieval • 5B • Updated • 266 • 4 -
vultr/VultronRetrieverFlash-Qwen3.5-0.8B
Visual Document Retrieval • 0.9B • Updated • 2.76k • 6
-
athrael-soju/colqwen3.5-4.5B-v3
Visual Document Retrieval • 5B • Updated • 5k • 17 -
athrael-soju/HydraQwen3.5-0.8B
Visual Document Retrieval • Updated • 14 -
athrael-soju/ColGemma4-E4B-IT-Base
Visual Document Retrieval • Updated • 1 -
athrael-soju/ColGemma4-E2B-IT-Base
Visual Document Retrieval • Updated • 1
Dual-head VLM: ColBERT retrieval + autoregressive generation by toggling one LoRA. Canonical 4B + 0.8B, omni proof-of-concept, baselines.
-
athrael-soju/HydraQwen3.5-4B
Visual Document Retrieval • Updated • 12 -
athrael-soju/HydraQwen3.5-0.8B
Visual Document Retrieval • Updated • 14 -
athrael-soju/HydraQwen2.5-Omni-3B
Visual Document Retrieval • Updated • 2 -
athrael-soju/DualHead-GritLM-Qwen3.5-4B
Visual Document Retrieval • Updated • 17 • 1
Visual document retrieval models on Qwen3.5 backbone. ViDoRe v3 leaderboard competitors, 128-dim multi-vector.
ColBERT-style late-interaction visual document retrieval adapters built on Google Gemma-4 (E2B and E4B variants).
-
The Price of Anarchy in Disaggregated Inference
Paper • 2606.17081 • Published • 1 -
Spatially-Grounded Document Retrieval via Patch-to-Region Relevance Propagation
Paper • 2512.02660 • Published • 1 -
Architecture-Aware LLM Inference Optimization on AMD Instinct GPUs: A Comprehensive Benchmark and Deployment Study
Paper • 2603.10031 • Published -
Hydra: Unifying Document Retrieval and Generation in a Single Vision-Language Model
Paper • 2603.28554 • Published