Models

Lucid Train is local-first and model-agnostic. Run open-source models on your own hardware, bring an OpenRouter key for the frontier, or point it at a self-hosted endpoint.

Local · Ollama (free, offline)

/download <tag> installs Ollama automatically if needed, pulls the model, and tunes it for low memory pressure (flash attention, q8_0 KV cache, models unload between turns). The setup screen marks each model ✓ fits, ◐ low free RAM, or ✗ too big for your machine.

ModelNeedsNotes
gpt-oss:20b~16 GB RAMBest tool-calling that fits 16 GB · the default recommendation.
qwen3-coder:30b~32 GB RAMBest local agentic coder (RL-trained on SWE-Bench).
qwen3:4b~6 GB RAMGreat small all-rounder for modest machines.
deepseek-r1 / ERNIE 4.5variesStrong reasoning specialists (chat-first).
shell
/download gpt-oss:20b
/model gpt-oss:20b

Cloud · OpenRouter (bring your own key)

One key unlocks the whole catalog. Your key is stored locally and only ever sent to OpenRouter. ★ Lucid's choice is kimi-k2.7-code · a coding specialist that balances quality and price. On a budget, deepseek-v4-pro is the cheapest near-frontier option.

ModelContext$/1M in→outNotes
kimi-k2.7-code262K$0.75 → $3.50★ Lucid's choice · coding specialist
deepseek-v4-pro1M$0.44 → $0.87cheapest near-frontier, 80.6% SWE-bench
qwen3.7-max1M$1.25 → $3.75long-horizon agentic
kimi-k2.7 / k2.6262K~$0.75 → $3.50general Kimi line
glm / minimax200KvariesGLM 5.1 · MiniMax M2.7
gpt-5.1 / grok / gemini200K-1Mpremiumproprietary frontier
shell
/apikey
/model kimi-k2.7-code

Self-hosted (OpenAI-compatible)

Point Lucid Train at any vLLM, llama.cpp, SGLang, or TGI server:

shell
$ lucid-train --base-url http://localhost:8000/v1 -m my-served-model

Failover

List backup models and the client walks them on provider errors:

shell
$ export LLM_FALLBACK_MODELS=qwen/qwen3-coder,deepseek/deepseek-chat-v3.1