Open-weight LLM models — the catalog big chat apps don't carry
Every big chat app decides for you which model you're allowed to use — one lab, one alignment, one editorial line. LivePair's chat exposes a catalog instead: frontier models and open-weight models side by side, billed per token at synced rates. This is what's actually in the list and how it works.
Why the catalog matters
Different models carry different alignment choices. Open-weight families — DeepSeek, Qwen, Kimi, GLM, Llama-class — were tuned by different labs with different editorial decisions than the US frontier models. On creative writing, opinionated analysis, and unorthodox questions, they genuinely answer differently — which is the point of offering them.
The big apps don't carry these models not because they're worse, but because a single-app product needs a single editorial line. A model catalog doesn't.
What's in the list
Over 50 curated text models: the frontier labs' latest (the current flagship chat models) alongside the open-weight line (DeepSeek, Qwen, Kimi, GLM, Llama-family). New major-lab versions are admitted automatically by the catalog pipeline as providers publish them; dated snapshots and duplicates get rejected by the same filter, so the list stays current instead of bloated.
Model selection happens in the chat UI — a searchable picker, not a settings page — and each model bills on its own synced per-token rates. Provider cost plus a fixed markup, so the cheap models are actually cheap: flash-class calls run fractions of a cent per turn.
For developers
The same catalog is reachable over an OpenAI-compatible API: POST /v1/agent/chat with a standard {model, messages, max_tokens} body, billed per call in USDC on x402 with no account, or on an lp_ API key against prepaid credits. The full request/response contract is in the MCP docs — point an agent at /docs/mcp and it can onboard itself.
Input-side checks on the API are a deterministic layer — hard categories only — not an LLM re-judging your prompt on every call. Everything else is the model's own alignment, which is the entire point of a catalog.
Honest limits
A catalog isn't consequence-free: illegal categories get a 400 at the deterministic gate, accounts that probe the platform get banned, and each model's provider alignment still applies. What you get is the honest middle — models the single-model apps won't carry, at API price, under rules stated once instead of a moralizing wrapper.