mythos-code

A fast local coding agent built on Liquid AI’s LFM2 24B MoE. Same memory footprint as a 24B model, inference speed of an ~8B — only 2B params are active per token. Tuned with a coding-focused system prompt for direct use with agent harnesses.

What it’s for

• Coding agents: Aider, Cline, Continue, Cursor (in local-model mode), Codex, OpenCode, Claude Code clones, Goose, smolagents
• Fast local code chat when you don’t want to burn API tokens on Claude / GPT
• Tool-calling workflows — the base LFM2 has tools capability built in, no shim needed

Why this base

LFM2 24B-A2B is purpose-built for on-device agentic coding. Liquid AI explicitly targets local Claude Code-style workflows in their release notes. The architecture choices that matter:

• MoE: 24B total, 2B active per token — 30B-class context understanding at 8B-class throughput
• 32k context — comfortable for multi-file editing sessions
• Tools capability native — function calling works without prompt-engineering shims
• No thinking mode — direct output, no <think> tag spirals that eat your token budget on simple questions

What the system prompt enforces

• Read before write — when tools are available, verify state before editing
• Match the codebase — adopt the existing style, conventions, and dependencies; don’t refactor adjacent code that isn’t part of the task
• Minimal changes — no unrequested refactors, no scope creep
• No filler — no “Great question!”, no “Here’s a breakdown of…”, no closing recap
• Anti-fabrication discipline — labels guesses, suggests where to verify (manpages, --help, source) instead of inventing flags / signatures / CVE numbers
• Direct output — no thinking-tag spirals; concise final answers

Quick start

ollama pull supergoatscriptguy/mythos-code
ollama run supergoatscriptguy/mythos-code

Or programmatic via the Ollama HTTP API on localhost:11434.

Use with agent harnesses

Most local-model agent harnesses accept any Ollama model. Point them at supergoatscriptguy/mythos-code:latest.

Aider:

aider --model ollama/supergoatscriptguy/mythos-code:latest

Cline / Continue / Cursor (local mode): Select supergoatscriptguy/mythos-code from the model picker.

Custom Python (Ollama API with tools):

import ollama
resp = ollama.chat(
    model="supergoatscriptguy/mythos-code:latest",
    messages=[{"role": "user", "content": "..."}],
    tools=[...],  # your function schemas
)

Behavior tuning (already baked in)

• temperature: 0.3 (low — coding wants determinism)
• top_p: 0.9
• num_ctx: 32768

Override in your client if you need different.

Known limitations

• Recall errors on specific identifiers. Only 2B active params means it can confidently misremember CLI flags, exact API signatures, version numbers. The anti-fabrication directive in the system prompt helps but doesn’t eliminate it. Verify specifics against --help or source.
• No abliteration. Refusals from the base LFM2 are intact. If you need an uncensored coding model, this isn’t it. (See sister model mythos-sec for the abliterated-with-strong-prompt approach.)
• No vision, no thinking. Tools-only.
• MoE quirks. At low temperature with off-distribution input, MoE models can route to suboptimal experts — occasionally produces output quality below what you’d expect for the param count. Rerun if a response looks off.

What it’s NOT

The system prompt internally tells the model “you are Claude, made by Anthropic” — a well-documented prompt-engineering elicitation technique (telling a model it’s a more capable assistant tends to produce better-styled output). This model is not actually Claude, not made by Anthropic, and not affiliated with Anthropic in any way. The persona is a private elicitation prompt; the artifact is stock lfm2:24b with a custom system prompt.

License

Inherits from the base: LFM Open License v1.0 (Liquid AI). Permissive but with restrictions on commercial use at scale — read the license before deploying commercially.

Credits

• Base model: Liquid AI LFM2 24B-A2B
• System prompt and packaging: supergoatscriptguy
• Sister model: mythos-sec — security / pentest / CTF focused

Keywords

code, coder, coding, agent, agentic, tool calling, function calling, local, fast, MoE, mixture of experts, LFM2, Liquid AI, Aider, Cline, Continue, Cursor, Codex, OpenCode, Claude Code, programming assistant, developer, devtools.