██████╗ ██████╗ ██████╗ ███████╗ ██╗███████╗ ██████╗ ██████╗ ███████╗████████╗██████╗ ██╗ ██╗
██╔════╝██╔═══██╗██╔══██╗██╔════╝ ██║██╔════╝ ██╔══██╗██╔═══██╗██╔════╝╚══██╔══╝██╔══██╗╚██╗ ██╔╝
██║ ██║ ██║██║ ██║█████╗ ██║███████╗ ██████╔╝██║ ██║█████╗ ██║ ██████╔╝ ╚████╔╝
██║ ██║ ██║██║ ██║██╔══╝ ██║╚════██║ ██╔═══╝ ██║ ██║██╔══╝ ██║ ██╔══██╗ ╚██╔╝
╚██████╗╚██████╔╝██████╔╝███████╗ ██║███████║ ██║ ╚██████╔╝███████╗ ██║ ██║ ██║ ██║
╚═════╝ ╚═════╝ ╚═════╝ ╚══════╝ ╚═╝╚══════╝ ╚═╝ ╚═════╝ ╚══════╝ ╚═╝ ╚═╝ ╚═╝ ╚═╝
struct Developer: Identifiable {
let id = UUID()
let name = "Aurélien Amsellem"
let handle = "@aamsellem"
let location = "Paris, France 🇫🇷"
let role = "Software Engineer & AI Toolmaker"
var passions: [String] {
["macOS native apps", "AI-augmented workflows", "Privacy-first design", "Gamification"]
}
var currentQuest: String {
"Building companions that make devs 10x — without selling their soul (or data)"
}
var philosophy: String {
"100% local. Zero telemetry. Your machine, your data, your rules."
}
}|
Compagnon virtuel gamifié pour macOS
|
L'assistant IA qui vous connaît vraiment
|
|
LLM models for the Olares Store
|
27 optimized AI apps for the Olares One
|
languages:
daily: [ Swift, TypeScript, Python ]
familiar: [ Rust, Go, Shell ]
apple_ecosystem:
ui: [ SwiftUI, AppKit, UIKit ]
tools: [ XcodeGen, Sparkle, SPM ]
targets: [ macOS, iOS ]
ai_tooling:
runtime: Claude Code (shell process, not API)
pattern: "enriched prompts + local memory + personality layer"
belief: "AI should amplify humans, not replace them"
selfhosted_ai:
platform: Olares One (RTX 5090M 24GB + 96GB DDR5, sm_120 Blackwell)
backends: [ llama.cpp, vLLM, vLLM-Omni, ExLlamaV3, custom forks (buun, am17an, Genesis) ]
models: [ Qwen3.6-27B-Dense, Gemma-4-26B-A4B, Gemma-4-E4B, Nemotron-3-Nano-30B, Qwen3.5-35B-A3B, Devstral-24B, GLM-4.7, Voxtral, OmniVoice ]
speeds: { Gemma4-DFlash: "214 t/s", Nemotron: "184 t/s", Qwen3.5-Vision: "131 t/s", Qwen3.5: "129 t/s", Gemma4-MTP: "119 t/s", Qwen3.6-Turbo: "88 t/s", Qwen3.6-LongCtx-262K: "77 t/s" }
features: [ MTP speculative decoding, DFlash speculative decoding, TurboQuant K8V4 KV, Unsloth Dynamic quants, native vision, ASR + TTS ]
apps: 27
format: Helm charts + Cloudflare Worker market
infrastructure:
hosting: GitHub Releases (DMG)
secrets: macOS Keychain
storage: Local Markdown files # no database, no cloud
updates: Sparkle 2 (EdDSA signed)
principles:
- "Privacy is not a feature, it's a foundation"
- "Ship native, not Electron"
- "Gamify the boring, automate the tedious"
- "If it leaves your machine, you should know about it"
[2024-01-15] Realized Claude Code + shell process = infinite local AI power
[2024-06-20] ULY v1.0 — persistent memory changes everything
[2025-01-10] Started Mochi Mochi — what if your AI companion was a cute rice ball?
[2025-03-xx] Mochi Mochi ships with gamification, Notion sync, 8 personalities
[2026-03-04] Still shipping. Still local. Still private. 🍡
[2026-03-07] olares-market — bringing open LLMs to the Olares Store 🧠
[2026-03-08] olares-one-market — custom market source for Olares One ⚡
[2026-03-13] Qwen3.5 129 t/s, GLM-4.7 131 t/s, TTS with voice cloning — all on 24GB 🔥
[2026-03-14] Nemotron 3 Nano 30B-A3B — 184 t/s on Olares One. New speed king 👑
[2026-04-01] TurboQuant rotation lands in llama.cpp — q4_0 KV cache = 2x context, same quality
[2026-04-05] Gemma 4 26B-A4B — 119 t/s with native vision. 16 apps on the market now 🚀
[2026-04-05] Voxtral ASR + Realtime + TTS — complete voice pipeline on Olares One 🎙️
[2026-04-10] llama.cpp b8740 — CUDA fused multiply for MoE, 19 apps on the market 🔧
[2026-04-30] vLLM v0.20.0 + TurboQuant K8V4 + MTP n=3 — Qwen 3.6 27B at 88 t/s on 24 GB Blackwell 💪
[2026-05-04] First public DFlash bench on consumer 24 GB Blackwell — buun-llama-cpp + spiritbuun's drafter, ~80 t/s ⚡
[2026-05-07] Gemma 4 E4B MTP, first Blackwell consumer mobile bench: 178 t/s, 77% draft acceptance 🎯
[2026-05-09] Qwen 3.6 27B Long Context — 77 t/s @ FULL 262K with havenoammo's UD-Q3_K_XL + am17an MTP 🚀
[2026-05-09] Gemma 4 26B-A4B + DFlash on vLLM tokenspeed-preview — 214 t/s. New speed king 👑
while alive { code(); ship(); iterate() }