CoolFace
Modelpublic

ram1234598766/Cesium2-vision-GGUF

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face
1likes129downloads
Model Card

Cesium2 Vision — GGUF (Q4KM + mmproj)

MORPH-AI v7 (Cesium2) Vision — a QLoRA fine-tune of Qwen2.5-VL-3B-Instruct (document VQA from The Cauldron + the full Cesium2 2026-knowledge/persona text set), merged and exported for llama.cpp/Ollama.

Files

FileRoleSize
cesium2-vision-q4_k_m.ggufLLM tower (fine-tuned)1.93 GB
cesium2-vision-mmproj-f16.ggufCLIP vision projector1.34 GB

Both files are required for image input.

Quick start (Ollama)

bash
ollama pull ram1234598766/cesium2-vision

Manual import: put both files in one folder with a Modelfile:

dockerfile
FROM ./cesium2-vision-q4_k_m.gguf
FROM ./cesium2-vision-mmproj-f16.gguf
PARAMETER num_ctx 8192
PARAMETER stop "<|im_end|>"

Model facts

BaseQwen/Qwen2.5-VL-3B-Instruct
Fine-tuneQLoRA r=16 on attention projections, 1 epoch
Data400 docvqa images + Cesium2 2026 persona/knowledge set
Architectureqwen2vl + clip projector

Links

⚡ What's new (August 2026)

  • —🌐 Web-connected Playground — toggle Live web search (keyless DuckDuckGo) and URL reading/scraping: grounded answers with cited sources
  • —💾 Sessions stored per account — full sessions API (GET/POST /v1/sessions), chat history follows you across devices
  • —🔑 Two key types — cs2-pg-… Playground keys (browser chat) and cs2-sk-… Agent keys (Claude Code, OpenAI SDK, curl), each with its own usage analytics
  • —🧰 Tools API — POST /v1/tools/search (live web), POST /v1/tools/read (scrape), plus Wikipedia, GitHub, Hugging Face, arXiv, exact-math and clock connectors via POST /v1/tools/use
  • —⌨️ VS Code extension 1.1.0 — inline completions, web-grounded chat, cloud sessions, 14 coding commands, slick animations
  • —⚡ Faster serving — models pre-warmed and pinned in RAM, consistent 8k context, 60s search cache, 8-way parallel
  • —🌐 Live API & key portal: <https://ram1234598766-soverign-cognitive-quantum-space.hf.space> (sign in with Google → free personal key, works with Claude Code)
  • —💻 Source & pipeline: <https://github.com/ram1234598766-dotcom/cesium2>
  • —📝 Text sibling model: <https://huggingface.co/ram1234598766/Cesium2-v7-GGUF>
  • —🧩 VS Code assistant with drag-&-drop vision: <https://open-vsx.org/extension/ram1234598766/cesium2-ai>
  • —🫥 Ollama page: <https://ollama.com/ram1234598766/cesium2-vision>

Author: Mrityunjay K — trained, fine-tuned, quantized and shipped end-to-end on free Kaggle GPUs. Live API: <https://ram1234598766-soverign-cognitive-quantum-space.hf.space>