Home

Camelid runs supported GGUF language and vision models locally with a Rust-native engine. Desktop app, browser chat, terminal UI, and an OpenAI-compatible API — all backed by the same single binary. No Python, Node.js, or Docker at runtime.

Camelid's tokenizer, GGUF loader, CPU kernels, Metal path, and CUDA path are all implemented in this repository. Model inference runs on your hardware.

What makes Camelid unusual is the second half of the sentence: it supports exact model files at exact quantizations, each validated token-for-token against a pinned llama.cpp reference. A file outside the validated set fails closed with a typed error instead of quietly running an unverified path.

Install

Windows 10/11 (x86_64) — signed per-user install, CUDA runtime bundled:

irm https://raw.githubusercontent.com/timtoole02/Camelid/main/scripts/get-desktop-windows.ps1 | iex

macOS 12+ (Apple Silicon) — installs to /Applications; re-run to update:

curl -fsSL https://raw.githubusercontent.com/timtoole02/Camelid/main/scripts/get-desktop-macos.sh | bash

Prefer the command line, or on Linux? Grab an engine archive from the latest release and:

camelid pull 3b_instruct_q8
camelid serve --model models/Llama-3.2-3B-Instruct-Q8_0.gguf

Full detail in Quick Start and Installation & Platforms.

Ways to use it

Interface	Start it with	Best for
Desktop app	Install from above	Native app, bundled engine, Models page
Browser chat	camelid serve --model <gguf>	Everyday local chat at 127.0.0.1:8181
Terminal UI	camelid chat	Shell and SSH workflows
HTTP API	camelid serve	OpenAI-compatible integration
Agent mode	camelid chat --agent --model <gguf>	Approval-gated tools in a repository
Workspace (preview)	Workspace in the web UI	Read-only analysis of a local folder

See Interfaces for what each one does and where its boundaries are.

Platform and acceleration

Platform	Distribution	Acceleration
Windows x86_64	Desktop installer, portable app, engine archive	CUDA on validated paths; CPU fallback
macOS Apple Silicon	Desktop DMG, engine archive	Metal and CPU
Linux x86_64	Engine archive	CUDA compiled in; CPU fallback

Acceleration claims are row- and configuration-specific — a GPU lane is validated per model row, per recorded GPU. See Compatibility & Evidence before relying on a particular GPU/model/quant combination.

Models

camelid pull ships a curated catalog of around thirty hash-pinned GGUFs spanning llama, qwen3, qwen25, qwen35, gemma3, gemma4, phi3, and command-r architectures, at quantizations from Q8_0 down through K-quants, i-quants, and PrismML ternary formats.

Good starting points:

Goal	Model	Pull ID
Smallest end-to-end test (~1.2 GB)	TinyLlama 1.1B Chat Q8_0	tinyllama
Recommended first model	Llama 3.2 3B Instruct Q8_0	3b_instruct_q8
Fits a 16 GB Apple Silicon Mac	Mistral 7B Instruct v0.3 Q8_0	mistral
Reasoning and coding on a small budget	Qwen3 4B Q4_K_M	qwen3_4b_q4
Compact PrismML GPU model	Bonsai 4B Q1_0	bonsai_4b_q1
Browser/API vision	Bonsai 27B Q1_0	bonsai_27b_q1

The full table, plus how pull IDs resolve, is in Model Catalog.

Important

The catalog and the support ledger are different lists. Everything in the catalog downloads and loads. A smaller set carries a validated exact-row support claim. Model Catalog marks which is which; Compatibility & Evidence is the contract.

Two reading paths

I want to use Camelid

1. Quick Start — install, pull a model, first token.
2. Installation & Platforms — distributions, acceleration, building from source.
3. Model Catalog — what to download and how big it is.
4. Interfaces — desktop, browser, terminal, agent, workspace.
5. API Reference — the OpenAI-compatible surface and everything else on the port.
6. Configuration & Deployment — flags, env vars, API keys, TLS, remote serving.

I want to audit Camelid's claims

1. Compatibility & Evidence — what "exact-row support" means and the live ledger.
2. Reproducing Parity Audits — run the same harnesses the project gates on.
3. Architecture — module boundaries, execution lanes, the inference path.
4. Roadmap & Promotion Criteria — what a row must satisfy before its claim moves.
5. Contributing & Validation — the validation gate and the naming/credit rules.

Source-of-truth docs

This wiki mirrors the repo's docs; it never forks them. Where content overlaps COMPATIBILITY.md, this wiki summarizes and links so it cannot drift. If this wiki and COMPATIBILITY.md ever disagree, COMPATIBILITY.md wins.

• README · COMPATIBILITY.md (authoritative) · STATUS.md · ROADMAP.md
• DOCS.md (documentation index) · CONFIGURATION.md · ARCHITECTURE.md
• RECEIPTS.md · BENCHMARKS.md · SUPPORT_MATRIX_v0.1.md

Important

Naming: the crate, binary, and command are all camelid (env vars CAMELID_*). The legacy name backendinference is forbidden by the repo's public-scrub CI — see decision D2.

Note

Credit: Camelid's tokenizer, reference compatibility layouts, and parity baselines are checked against llama.cpp (© 2023–2026 The ggml authors, MIT). Camelid keeps its own Rust-native codebase while crediting the ggml ecosystem.