Developer Gives DeepSeek V4 Vision for Just $2,000 by Training a 40M Connector
ButtercupLyn100 · reddit · 2026-08-11
A developer successfully added basic vision capabilities to the text-only MoE model DeepSeek V4 Flash without modifying the underlying language model.
Experiment Design:
- Froze the 284B-parameter DeepSeek V4 Flash and the 417M-parameter MoonViT image encoder.
- Trained only a 40.1M-parameter connector (MLP) between them.
- Trained for 1 epoch on 100,000 image-text examples sampled from The Cauldron dataset.
Training & Cost:
- Used 5× H200 GPUs with an effective throughput of 4,938 examples/hour.
- Total end-to-end project cost was approximately $2,000, including data prep, conversion, and inference validation.
Results:
- The model gained functional vision, successfully recognizing objects in street photos and identifying specific UI controls in browser screenshots.
- While still an experimental pilot, it proves the viability of retrofitting large text models with vision via a lightweight, low-cost connector.
More from Models
- Local Open-Source Model Self-Checks and Runs Code: What Do Subscriptions Still Buy? — truecakesnake · 2026-08-11
- OpenAI Accused of Stifling Model Creativity via Strict Temperature Controls — RileyRalmuto · 2026-08-11
- GLM-5.2 Pricing Drops: High-Level Intelligence Gets Dramatically Cheaper — tobowers · 2026-08-11
- Developer Praises Kimi K3: 'Like if Llama 4 Behemoth Was Real' — willccbb · 2026-08-11
- Alibaba Confirms Qwen3.8-27B Open Weights Landing This Week — max_paperclips · 2026-08-11
- Abacus AI Releases Smaug-Agentic, Topping Open-Source Leaderboard for Agentic Coding — bindureddy · 2026-08-11