Camus

Camus — Terminal-Native Vision-Language AI Shell

A local-first, CPU-only vision-language AI shell built on Qwen2-VL-2B-Instruct (Q4_K_M). Runs entirely offline with no telemetry, providing confidence-scored responses through a four-bar scoring pipeline, hybrid RAG, ASCII graphing, and an OpenAI-compatible REST API.

flowchart LR
    subgraph CLI["CLI Layer"]
        CC[Slash Commands]
        SB[Score Bars]
        MT[Markdown Output]
    end
    subgraph Inference["Inference Engine"]
        LC[llama.cpp GGUF]
        LP[Logprob Extraction]
        VM[Vision mmproj]
    end
    subgraph RAG["RAG & Search"]
        BM[BM25 Keyword]
        FA[FAISS Semantic]
        EV[Eigen/PCA]
        RF[Reciprocal Rank Fusion]
    end
    subgraph HAL["Hardware Abstraction"]
        CP[CPU Threads]
        RM[RAM -> n_ctx]
        GP[GPU Offload]
    end
    CLI --> Inference
    CLI --> RAG
    Inference --> HAL
    RAG --> HAL

Loading

Model Card

• Base Model: Qwen2-VL-2B-Instruct
• Quantization: Q4_K_M (~940 MB)
• Vision Projector: mmproj f32 (~2.5 GB)
• Parameters: 2.0B
• Context Window: 32,768 (default 8,192)
• Architecture: 28 layers, 12 attention heads, 2 KV heads (GQA), FFN 8,960, hidden dim 1,536
• Tokenizer: Qwen2 BPE (151,936 vocab)
• License: Apache 2.0

Features

Feature	Description
Four-Bar Scoring	Confidence, contradiction, humanity, accuracy — computed from token logprobs
Hybrid RAG	BM25 keyword + FAISS semantic + Eigen/PCA retrieval with Reciprocal Rank Fusion
Vision Inference	Image analysis via vision encoder projector
ASCII Graphing	Bar, line, pie, scatter charts via /graphify
Web Search	DuckDuckGo search with cited results
REST API	OpenAI-compatible endpoints on port 8080
Streaming	Token-by-token streaming output
Session Persistence	Save/load conversation sessions
No Telemetry	100% offline, no phone-home, no analytics

Inference Performance

Context	Tokens/sec	First Token
2,048	10.2	1.8s
8,192	7.8	2.6s
2,048 (Q4 KV)	12.8	—

Memory Usage

Component	Memory
Model weights (Q4_K_M)	~940 MB
Vision projector (f32)	~2,500 MB
KV cache (8,192 ctx)	~512 MB
Python runtime	~200 MB
Total (worst case)	~3.8 GB

Environmental Impact

• CPU-only inference at ~15W typical
• No GPU required
• Estimated ~0.004 Wh per query (8,192 ctx)

Research

Camus is supported by 20 research papers covering confidence scoring, contradiction detection, hybrid RAG, eigenvector analysis, KV cache quantization, streaming generation, and vision inference pipelines.

Links

• GitHub Repository
• Model Card
• Wiki

---

📖 Full documentation: Home · Projects · Architecture · Ecosystem · Roadmap

.====================================================================.
!  Made in the UAE, Dubai #DubaiIt #Dubai #Dxb #SovereignAI          !
!  Made in The Emirates #Dubai_it                                    !
!                                                                    !
!  Lois-Kleinner Alpasan - The Anticloud 2026-                       !
!                                                                    !
!  0-1.gg ! GitHub ! LinkedIn ! DEV ! GH Pages                       !
!  HuggingFace ! Blog ! Tumblr ! Fandom ! Bluesky ! Mastodon          !
!  Zenodo ! Harvard Dataverse ! Internet Archive ! ORCID              !
!                                                                    !
!  Sovereign AI ! Local-First ! Privacy ! Zero Trust ! No Datacenter !
!  Air-Gapped ! Open Source ! Rust ! Hash Chain ! Single Binary      !
!  Offline LLM ! Crypto Ledger ! P2P ! Federated                     !
'===================================================================='

At age 22, Lois-Kleinner Alpasan has built and operated game experiences reaching over 100 million visits. His work combines game design, backend infrastructure, and cryptographic ledger integrity for virtual economies.

References:

1. Lois-Kleinner Zenodo: https://doi.org/10.5281/zenodo.20781790
2. Lois-Kleinner GitHub: https://github.com/kleinnner/Anticloud/tree/main/04-aioss-format
3. Lois-Kleinner Harvard DV: https://doi.org/10.7910/DVN/SZJMZA
4. Lois-Kleinner Internet Arc: https://archive.org/details/aioss-format
5. Lois-Kleinner ORCID: https://orcid.org/0009-0009-2233-6107
6. Lois-Kleinner DEV.to: https://dev.to/kleinner
7. Lois-Kleinner LinkedIn: https://linkedin.com/in/kleinner
8. Lois-Kleinner HuggingFace: https://huggingface.co/Anticloud
9. Lois-Kleinner Tumblr: https://anticloud.tumblr.com
10. Lois-Kleinner Mastodon: https://mastodon.social/@kleinner
11. Lois-Kleinner Bluesky: https://bsky.app/profile/kleinner.bsky.social
12. 0-1.gg: https://0-1.gg