webbrain-one/DeepSeek-V4-Flash-0731-Vision-NVFP4 路 Hugging Face
Log In<br>Sign Up
","lstrip":false,"normalized":true,"rstrip":false,"single_word":false},"eos_token":{"__type":"AddedToken","content":"","lstrip":false,"normalized":true,"rstrip":false,"single_word":false},"pad_token":{"__type":"AddedToken","content":"","lstrip":false,"normalized":true,"rstrip":false,"single_word":false},"unk_token":null}},"createdAt":"2026-08-04T19:02:09.000Z","discussionsDisabled":false,"discussionsSorting":"recently-created","downloads":0,"downloadsAllTime":0,"id":"webbrain-one/DeepSeek-V4-Flash-0731-Vision-NVFP4","isLikedByUser":false,"availableInferenceProviders":[],"showHuggingChatEntry":false,"inference":"","lastModified":"2026-08-05T06:01:58.000Z","likes":1,"pipeline_tag":"image-text-to-text","library_name":"custom","librariesOther":[],"trackDownloads":true,"model-index":null,"private":false,"repoType":"model","gated":false,"tags":["custom","safetensors","deepseek_v4","multimodal","vision-language","deepseek-v4","moonvit","nvfp4","fp8","blackwell","sglang","image-text-to-text","base_model:MJPansa/DeepSeek-V4-Flash-0731-NVFP4","base_model:quantized:MJPansa/DeepSeek-V4-Flash-0731-NVFP4","license:other","8-bit","region:us"],"tag_objs":[{"id":"image-text-to-text","label":"Image-Text-to-Text","type":"pipeline_tag","subType":"multimodal"},{"id":"safetensors","label":"Safetensors","type":"library"},{"id":"custom","label":"custom","type":"other","clickable":true},{"id":"deepseek_v4","label":"deepseek_v4","type":"other","clickable":true},{"id":"multimodal","label":"multimodal","type":"other","clickable":true},{"id":"vision-language","label":"vision-language","type":"other","clickable":true},{"id":"deepseek-v4","label":"deepseek-v4","type":"other","clickable":true},{"id":"moonvit","label":"moonvit","type":"other","clickable":true},{"id":"nvfp4","label":"nvfp4","type":"other","clickable":true},{"id":"fp8","label":"fp8","type":"other","clickable":true},{"id":"blackwell","label":"blackwell","type":"other","clickable":true},{"id":"sglang","label":"sglang","type":"other","clickable":true},{"id":"base_model:MJPansa/DeepSeek-V4-Flash-0731-NVFP4","label":"base_model:MJPansa/DeepSeek-V4-Flash-0731-NVFP4","type":"other","clickable":true},{"id":"base_model:quantized:MJPansa/DeepSeek-V4-Flash-0731-NVFP4","label":"base_model:quantized:MJPansa/DeepSeek-V4-Flash-0731-NVFP4","type":"other","clickable":true},{"id":"8-bit","label":"8-bit precision","type":"other","clickable":true},{"id":"license:other","label":"other","type":"license"},{"type":"region","label":"馃嚭馃嚫 Region: US","id":"region:us"}],"safetensors":{"parameters":{"BF16":1940552560,"F32":37741630,"I64":2327040,"F8_E4M3":6304038912,"U8":277025390592,"I8":19327352832},"total":304637403566,"sharded":true,"totalFileSize":176476253712},"hasBlockedOids":false,"region":"us","isQuantized":true,"licenseFilePath":"LICENSE"},"discussionsStats":{"closed":0,"open":0,"total":0},"query":{},"inferenceContextData":{"billableEntities":[],"entityName2Providers":{}},"hasQuantizations":false,"copyToBucketNamespaces":[]}">
DeepSeek V4 Flash 0731 Vision (NVFP4)
DeepSeek V4 Flash 0731 with sight. This development checkpoint connects<br>DeepSeek's reasoning and agentic backbone to the MoonViT vision encoder from<br>Kimi-K2.6 through WebBrain's<br>trained, routing-aware PatchMerger projector.
This is a complete copy of MJPansa's source-preserving 0731 NVFP4 conversion<br>with WebBrain's frozen MoonViT tower, trained PatchMerger projector, routing<br>bridge, and custom SGLang integration added in the same repository.
The text checkpoint is mixed precision: routed expert weights are NVFP4 while<br>attention, shared experts, head, MTP, and other ignored paths preserve the<br>source 0731 formats described by config.json, hf_quant_config.json, and<br>conversion-receipt.json. It is not an all-NVFP4 checkpoint.
Why vision at WebBrain
At WebBrain, we build browser agents that need to<br>understand the visual state of the web鈥攏ot just extracted text. Screenshots,<br>charts, dashboards, rich editors, and the location and appearance of controls<br>are part of real browser work, so vision is a practical product requirement.
In our<br>American鈥揅hinese open-model frontier benchmark,<br>DeepSeek V4 Flash stood out as a strong and cost-efficient model, while the<br>upstream checkpoint remains text-only. This project attaches a MoonViT vision<br>bridge while keeping the language backbone and vision tower frozen.
Contents
Component<br>Provenance<br>Status
Text checkpoint<br>MJPansa/DeepSeek-V4-Flash-0731-NVFP4 at 64d64cd89bc63a66aa46506da89d7821f7491c62<br>48 shards; copied server-side without tensor changes
Conversion source<br>deepseek-ai/DeepSeek-V4-Flash-0731 at 9e165c30e2704aec5d9d593cce3eebd58bbef1cb<br>0731 DSpark/MTP tensors preserved
Vision tower<br>MoonViT-3d extracted from moonshotai/Kimi-K2.6 at 7eb5002f6aadc958aed6a9177b7ed26bb94011bb<br>329 BF16 tensors; frozen
Projector<br>WebBrain PatchMerger, trained with the text backbone and tower frozen<br>6 BF16 tensors; 40,119,040...