Local-first AI models from Mithraeum. Quiet weights for people who write code on their own machine.

_{site  ·  releases  ·  hako-code  ·  hako-edit  ·  org}

hako-sho-stock as the default in hako-code	:models · live stock-wraps + queued fine-tunes
a hako model inside hake (the editor)	shared mithraeum palette across the suite

hakm CLI captures pending. hakm run, hakm list, hakm pull shots will replace the reused suite images above as they're recorded. The agent + editor stand-ins demonstrate the live stock-wraps (hako-sho-stock, hako-koi-mini-stock) working end-to-end today.

hako is a growing model family.

╭───────────────────────────────────────────╮
│        ⠀⠀⠀⠀⠀⠀⠀⢀⣀⡀⠀⠀⠀⠀⠀⠀⢀⣀⡀⠀⠀⠀⠀⠀⠀⠀         │
│        ⠀⠀⠀⢀⣠⣶⣾⣿⣿⣿⣦⡀⠀⠀⢀⣴⣿⣿⣿⣷⣶⣄⡀⠀⠀⠀         │
│        ⣠⣴⣾⣿⣿⣿⣿⣿⣿⠿⠛⠉⢠⡄⠉⠛⠿⣿⣿⣿⣿⣿⣿⣷⣦⣄         │
│        ⠀⠻⣿⣿⣿⡿⠟⠋⠁⠀⠀⠀⢸⡇⠀⠀⠀⠈⠙⠻⢿⣿⣿⣿⠟⠁         │
│        ⠀⠀⠈⠋⠁⠀⠀⠀⠀⠀⠀⠀⢸⡇⠀⠀⠀⠀⠀⠀⠀⠈⠙⠁⠀⠀         │
│        ⠀⠀⣰⣷⣦⣄⡀⠀⠀⠀⠀⠀⢸⡇⠀⠀⠀⠀⠀⢀⣠⣴⣾⣆⠀⠀         │
│        ⢠⣾⣿⣿⣿⣿⣿⣷⣦⣄⠀⠀⢸⡇⠀⠀⣠⣴⣾⣿⣿⣿⣿⣿⣷⡄         │
│        ⠙⠿⣿⣿⣿⣿⣿⣿⣿⣿⣿⠖⠀⠀⠲⣿⣿⣿⣿⣿⣿⣿⣿⣿⠿⠋         │
│        ⠀⠀⠀⠉⠛⠿⣿⣿⣿⠟⢁⣴⡇⢸⣦⡈⠻⣿⣿⣿⠿⠛⠉⠀⠀⠀         │
│        ⠀⠀⠀⢸⣷⣦⣄⡉⢁⣴⣿⣿⡇⢸⣿⣿⣦⡈⢉⣠⣴⣾⡇⠀⠀⠀         │
│        ⠀⠀⠀⠈⠙⠻⢿⣿⣿⣿⣿⣿⡇⢸⣿⣿⣿⣿⣿⡿⠟⠋⠁⠀⠀⠀         │
│        ⠀⠀⠀⠀⠀⠀⠀⠈⠙⠻⢿⣿⡇⢸⣿⡿⠟⠋⠁⠀⠀⠀⠀⠀⠀⠀         │
│        ⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠈⠁⠈⠁⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀⠀      
│                                           │
│ hako 0.1.6 · mithraeum · hako-sho-stock   │
│ trust on · session resumed                │
│ :help :providers :models :login :theme    │
╰───────────────────────────────────────────╯

 ▄█████▄
██ ███ ██ sho       : mini       · Qwen2.5-Coder-3B  · stock-wrap live
█████████ koi-mini  : mid-small  · Qwen2.5-Coder-7B  · stock-wrap live
█████████ koi       : mid        · 14B / 32B target  · fine-tune queued
▀█▀▀█▀▀█▀ samurai   : max        · 50B+ reserved     · waits on hardware

_{—— I ——}

Overview

• Local first. Models run on your device. Offline. No cloud inference. No API calls at runtime.
• Four tiers. mini (sho, 3B, phone-reachable), mid-small (koi-mini, 7B, M-class iPad and laptop), mid (koi, 14B/32B target, desktop), max (50B+ reserved). Quantization is the brand promise: older hardware stays useful.
• Honest framing. Stock-wrap is labeled stock-wrap. Fine-tune is labeled fine-tune. Every Modelfile SYSTEM and model card says exactly what it is.
• One CLI: hakm. Standalone runner. hakm run sho "...", hakm chat koi-mini, hakm list, hakm pull sho. Shell stub today, native runner in v0.1.7.
• Wired into the suite. hako (the agent) auto-defaults to whichever hako-*-v* is installed (preference: koi > koi-mini > sho). hake (the editor) embeds the agent in a split pane. Zero config when all three are present.
• Tool-use ready. Each Modelfile SYSTEM teaches the <tool name="...">{...}</tool> schema the agent parses. Works on a 3B model, not just chat.
• ChatML throughout. <|im_start|> / <|im_end|> template. No vendor lock.
• Naming. Stock-wraps carry no version (hako-sho-stock). The first real fine-tune of a tier earns v0.0.1 (hako-koi-v0.0.1). hakm resolves shorthand: sho → hako-sho-stock, koi-mini → hako-koi-mini-stock; koi and samurai are queued.

_{—— II ——}

Install

One-line install (recommended)

curl -fsSL https://mithraeums.github.io/hakm.sh | sh

Installs hakm + pulls the live stock-wraps: hako-sho-stock (3B) and hako-koi-mini-stock (7B). Run hakm run sho "hello" immediately after.

v0.1.6 note: transport is ollama today (transitional). The native hakm-server lands v0.1.7 and drops every external runtime dep.

Manual

# 1. clone
git clone https://github.com/mithraeums/hako.git && cd hako

# 2. install hakm into ~/.local/bin (or anywhere on PATH)
install -m 0755 hakm ~/.local/bin/hakm

# 3. build a stock-wrap (base + bundled Modelfile)
hakm pull sho               # or: hakm pull koi-mini

Verify

hakm --version              # hakm 0.1.0
hakm list                   # NAME                  ID   SIZE   MODIFIED
                            # hako-sho-stock        ...
hakm run sho "write fib"    # streams response

_{—— III ——}

Tiers

Tier	Project	Base	Today	Target	Runtime	Hardware
mini	sho	Qwen2.5-Coder-3B-Instruct	stock-wrap	3B fine-tune on mithraeum docs	ollama (hakm-server v0.1.7)	iPhone, iPad, iSh, anything CPU
mid-small	koi-mini	Qwen2.5-Coder-7B-Instruct	stock-wrap	7B fine-tune	ollama (hakm-server v0.1.7)	M-class iPad, any laptop
mid	koi	Qwen2.5-Coder-14B (or 24B)	not pulled	32B fine-tune on rented A100	ollama, then hakm-server	desktop, workstation
max	samurai	TBD	not pulled	50B (later 128B, 200B)	hakm-server	workstation, dGPU

Model catalog

Tag	Tier	What	Status
hako-sho-stock	mini	Qwen2.5-Coder-3B-Instruct + hako SYSTEM wrap	live
hako-koi-mini-stock	mid-small	Qwen2.5-Coder-7B-Instruct + hako SYSTEM wrap	live
hako-sho-v0.0.1	mini	3B fine-tune on mithraeum docs + curated code corpus	queued
hako-koi-mini-v0.0.1	mid-small	same recipe at 7B	queued
hako-koi-v0.0.1	mid	14B/32B base + LoRA fine-tune via rented A100	queued
hako-samurai-v0.0.1	max	50B+ base + fine-tune	reserved, hardware-blocked

Scratch-from-scratch research (byte-level GPT, own C17 inference, .mlf format) lives in experimental/ and is gitignored. Not part of the release pipeline.

_{—— IV ——}

Use

Standalone via hakm

hakm run sho "explain ring buffers in C"
hakm chat koi-mini               # interactive REPL
hakm list                        # installed hako-* models
hakm pull sho                    # download base + create hako-sho-stock
hakm models                      # full catalog

hakm <short> resolves shorthand to the current tag (sho → hako-sho-stock, koi-mini → hako-koi-mini-stock, koi → hako-koi-v0.0.1 when shipped).

Via hako-code (the agent)

If any installed hako-* model is in the local list, hako auto-defaults on first launch (preference order: koi > koi-mini > sho, fine-tunes over stock-wraps), no config:

hako                             # detects local hako model, defaults to mithraeum provider
›  list the files in this directory
◆  <tool name="list_dir">{"path": "."}</tool>
r list_dir(.)
  README.md  hakm  koi/  sho/  LICENSE
◆  README, the hakm CLI, two model dirs, and a LICENSE file.

Or set explicitly:

:provider mithraeum
:model hako-sho-stock           # or hako-koi-mini-stock / hako-koi-v0.0.1 when shipped

Via hake (the editor)

Open hake, drop into the AI pane (Ctrl-W r), set :provider mithraeum :model hako-sho-stock (or any installed hako-* model). Tool calls run in the editor's split pane against the file you're looking at.

_{—— V ——}

Layout

hako/                          (this repo, mithraeums/hako)
├─ hakm                        standalone CLI runner (shell stub today, C in v0.2)
├─ koi/                        mid + mid-small tiers, Qwen-based, mithraeum runtime
│   ├─ sft/                    LoRA fine-tune pipeline (axolotl)
│   ├─ models/                 Modelfiles + (gitignored) GGUF weights
│   │   └─ hako-koi-mini-stock/   7B stock-wrap, tool-aware SYSTEM
│   └─ data/                   fine-tune datasets (gitignored)
├─ sho/                        mini tier, 3B stock-wrap today
│   └─ models/
│       └─ hako-sho-stock/        3B stock-wrap, tool-aware SYSTEM
└─ experimental/               from-scratch research, gitignored, not shipping
    ├─ nano/                   byte-level GPT trainer (Python + torch)
    ├─ engine/                 C17 inference, libc + libm, mmap'd .mlf
    ├─ models/                 .mlf weight files
    └─ data/                   pretrain corpora

_{—— VI ——}

Design constraints (shared across tiers)

• ChatML template (<|im_start|> / <|im_end|>)
• Tool-call format compatible with hako-code's HK_TOOLS schema (prose <tool name="X">{...}</tool> parsed at the agent boundary)
• Local-only inference. No cloud at runtime.
• Honest framing in every Modelfile SYSTEM and model card. Stock is stock. Fine-tunes name their base and corpus.
• C-first runtime. Ollama is the transport today. hakm-server (v0.1.7) replaces it: minimal llama.cpp subset, CPU-only, drops the runtime dep. v0.2 swaps the vendored kernels for hand-tuned RoPE / RMSNorm / SwiGLU / attention / Q4_K dequant in AVX2 + NEON.

_{—— VII ——}

The suite

Product	CLI	Source	Role
Models (this)	hakm <model> <prompt>	mithraeums/hako	local weights + runner
Agent	hako	mithraeums/hako-code	terminal AI agent
Editor	hake	mithraeums/hako-edit	modal terminal editor (embeds the agent)

_{—— VIII ——}

Roadmap

• hako-sho-v0.0.1 first real fine-tune on mithraeum docs + curated code corpus, 3B base, rented A100.
• hako-koi-mini-v0.0.1 same recipe at 7B.
• hako-koi-v0.0.1 mid-tier debut: 14B/32B base, LoRA on rented A100. Tooling staged in koi/sft/.
• hako-samurai-v0.0.1 max tier, 50B+, reserved on hardware.
• Native engine (v0.1.7) hakm-server: vendored minimal llama.cpp subset for Qwen2 Q4_K_M, CPU-only. Drops ollama runtime dep.
• Own kernels (v0.2) RoPE / RMSNorm / SwiGLU / attention / Q4_K dequant, hand-tuned AVX2 + NEON. Drops every line of vendored code.
• mithraeum-native provider in hako-code today :provider mithraeum aliases to ollama transport; v0.1.7 swaps to hakm-server over a local port.
• samurai (max) waits on hardware (target: RTX 5090 + 128GB DDR5).
• Scale plan 2x-4x each tier when training budget allows. Eventual lineup roughly 6B sho, 14B-24B koi-mini, 64B-128B koi, 200B max.

_{—— IX ——}

License

• sho, koi-mini, koi: Apache 2.0 (matches Qwen2.5-Coder base)
• Top-level LICENSE covers shared assets (READMEs, configs, the hakm CLI)
• experimental/ (gitignored, scratch research): MIT when published, not shipping today

_{— deus sol invictus mithras —}