Name: TinyTitan (dsh-tinytitan)
Author: pummelchen

Installer

dsh plugin --profile web add github:pummelchen/tinytitan

Ce plugin se trouve dans le sous-répertoire plugins/dsh-tinytitan du dépôt.

README

TinyTitan

TinyTitan runs Qwen-family MoE and dense text models, and Poolside's Laguna-S-2.1, locally on Apple Silicon, streaming the routed experts from SSD so a model larger than your RAM still runs. Decode is native Swift over hand-written Metal kernels; the prompt runs on the GPU or, optionally, the Apple Neural Engine. It ships a CLI, an installer/repacker, and a loopback OpenAI-compatible server with no MLX and no GGUF dependency.

Who it is for. People on an M-series Mac who want to run a 35B — or a 122B/125B — model locally on a machine that could never hold it in RAM, and developers who want a loopback OpenAI- or Anthropic-compatible endpoint for Codex, Claude Code, Qwen Code, OpenCode, Zed or the official SDKs. It is a terminal-first tool: the engine and its server are the product, and the DeepSeek Harness plugins are an optional client of that server. It is not a fine-tuning toolkit, a vision model, a GUI app, or a way to expose a model to your network — the server is 127.0.0.1-only and has no authentication.

What this repository delivers

The LLM engine as a library — TinyTitanLib. Embed the engine in your own Swift program: an Engine and a Session, streaming tokens through a callback. Same kernels, format reader and sampler as the engine, in your process — no subprocess, no HTTP, no model reimplementation.
The complete local LLM engine — TinyTitanCLI + TinyTitanServer. Run a 35B, 122B or 125B model from the terminal and serve it on a loopback OpenAI-/Anthropic-compatible endpoint for Codex, Claude Code, Qwen Code, OpenCode, Zed or the official SDKs. Native Swift and Metal: no MLX, no GGUF.
The DeepSeek Harness plugin — plugins/dsh-tinytitan. The harness finds the models you have installed and keeps its route current, with a quiet compaction backend.
The .ssdai model format and its converter — TinyTitanRepack + tools/. Stream a checkpoint into a hash-verified install that keeps the routed experts on SSD, so a model larger than your RAM still runs — and a truncated, swapped or edited install is refused rather than served.

Those four ship from one branch (main) and one release tag. The first two are the products — a library and an engine — and the rest are the plugin and tooling that ship beside them; the split is spelled out under Library and engine.

Quick start

Requires macOS 26+, Xcode 27 (Swift 6.4) and Apple Silicon.

git clone https://github.com/Pummelchen/TinyTitan.git
cd TinyTitan
swift build -c release
.build/release/TinyTitanCLI --help

Install a model with tools/install_models.sh --choose (the dense 2B row is about 2.3 GB), then run a prompt:

.build/release/TinyTitanCLI --model models/qwen3.5_2B_4Bit \
  --prompt "What is the capital of Paris ?" --max-new 96

Start the server with tools/server_launcher.sh; it prints the base URL and writes the client configuration. The full walkthrough is Getting Started.

Benchmarks

Peak decode on a base 8-core M3 MacBook Pro with 24 GB, after a warm-up pass.

Model	Quantization	GPU	CPU	Terminal-Bench 2.x	Terminal-Bench 4.0
Qwen3.8-Flash-Next 125B-A6B	4-bit	5.5 tok/s	NA	86% (2.1)	25%
Qwen3.8-Flash-Next 125B-A6B	8-bit	2.1 tok/s	NA	86% (2.1)	25%
Ornith 1.5 122B	4-bit	2.9 tok/s	NA	81% (2.1)	NA
Ornith 1.5 122B	8-bit	1.2 tok/s	NA	81% (2.1)	NA
Laguna-S-2.1 118B-A8B	4-bit	1.5 tok/s	NA	70% (2.1)	NA
Laguna-S-2.1 118B-A8B	8-bit	0.7 tok/s	NA	70% (2.1)	NA
Ornith 1.5 35B-A3B	4-bit	21.7 tok/s	NA	68% (2.1)	NA
Ornith 1.5 35B-A3B	8-bit	11.9 tok/s	NA	68% (2.1)	NA
K2-Horizon-MoVA-36B-A4B	4-bit	12.4 tok/s	NA	59% (2.1)	NA
K2-Horizon-MoVA-36B-A4B	8-bit	2.0 tok/s	NA	59% (2.1)	NA
Qwen 3.6 35B-A3B	4-bit	21.4 tok/s	NA	53% (2.1)	NA
Qwen 3.6 35B-A3B	8-bit	12.4 tok/s	NA	53% (2.1)	NA
NVIDIA Nemotron 3.5 Bolt 74B-A9B	4-bit	6.7 tok/s	3.6 tok/s	51% (2.1)	NA
NVIDIA Nemotron 3.5 Bolt 74B-A9B	8-bit	0.6 tok/s	0.9 tok/s	51% (2.1)	NA
Qwen-AgentWorld 35B-A3B	4-bit	21.7 tok/s	NA	40% (2.0)	NA
Qwen-AgentWorld 35B-A3B	8-bit	12.3 tok/s	NA	40% (2.0)	NA
KAT-Coder-V2.5-Dev 35B-A3B	4-bit	17.9 tok/s	NA	33% (2.1)	NA
KAT-Coder-V2.5-Dev 35B-A3B	8-bit	6.9 tok/s	NA	33% (2.1)	NA
Qwen 3.5 4B (dense)	4-bit	22.8 tok/s	7.5 tok/s	26% (2.1)	NA
Qwen 3.5 4B (dense)	8-bit	14.1 tok/s	7.9 tok/s	26% (2.1)	NA
Qwen 3.5 9B (dense)	4-bit	12.9 tok/s	4.4 tok/s	9% (2.0)	NA
Qwen 3.5 9B (dense)	8-bit	7.8 tok/s	4.4 tok/s	9% (2.0)	NA
Qwen 3.5 2B (dense)	4-bit	44.4 tok/s	17.4 tok/s	3% (2.1)	NA
Qwen 3.5 2B (dense)	8-bit	29.6 tok/s	18.1 tok/s	3% (2.1)	NA

Documentation

Everything is in the Wiki.

Credits

TinyTitan is a focused fork of drumih/turbo-fieldfare, whose bounded-memory runtime, installer, CLI and local server this project builds on. The Qwen 3.6 integration was created by NeelM0906 in upstream PR #29. Concise mode is derived from the Nail-Qwen3.6-35B-A3B chat template by peculiar-ragdoll.

Project Maya: the 6.0 checkpoint, calibration and machine-size cache work was carried under that name. The live ledger is docs/maya-task-status.md; the closeout is docs/maya-closeout-2026-10-22.md.

Kapitsa: the cross-session shared system-prefix cache and the M4+ tensor-ops prefill direction were informed by nace-ai/kapitsa. No code is copied from it.

Related research: TinyTitan Datacenter runs large MoE models across a cluster of Mac minis and Studios, keeping them on SSD/NVMe for near-linear scale of decode throughput.