Chat

hipfire provides two native Rust chat surfaces:

• hipfire tui (or bare hipfire) opens the full ratatui application and its Chat tab.
• hipfire chat [model] is a lightweight line-oriented multi-turn client for terminals and scripts.

Both use the native OpenAI-compatible service documented in SERVE.md.

Full terminal UI

hipfire
# or
hipfire tui

Open the Chat tab, select a downloaded model in Models, or use /model <tag>. The TUI supports streaming, multiline input (Ctrl+O), cancellation, saved sessions, model/system/sampling commands, regeneration/edit/copy actions, and live serve status. Run /help in Chat for the current command list.

The TUI asks the native CLI to start a detached service when the configured endpoint is offline. Config, registry, model lifecycle, and HTTP calls all use the shared Rust crates; there is no separate script runtime.

Lightweight chat command

hipfire chat qwen3.6:35b-a3b-mq4r

you> Explain wave32 WMMA briefly.
assistant> ...
you> /clear
you> /exit

Supported local commands are /clear, /exit, and /quit. Flags:

If no healthy service exists, the command starts a tracked detached native service on the configured host/port with lazy model loading. That service remains available after the line client exits; use hipfire stop when you want to release it.

Resolution and request behavior

• Model names use the same registry/alias/path resolver as run and serve.
• The service resolves load-time KV/speculation/model overrides from config.toml, models.toml, registry recommendations, and compatible env.
• Explicit chat sampling flags are sent on every turn. Omitted sampling fields follow the service’s per-model/registry/daemon fallback contract.
• The full messages array is sent each turn, so conversation context is preserved by request content rather than hidden client state.
• Streaming answer and reasoning deltas share the same Rust HTTP client used by the TUI.

Thinking and model framing

Reasoning models may emit a thinking stream before the answer. The HTTP API exposes it as reasoning_content; visible answer text remains content.

Mode, semantic effort, and hard cap are independent — see CONFIG.md for keys/defaults and SERVE.md for per-request HTTP fields (reasoning_effort, enable_thinking, max_think_tokens, legacy thinking_budget). Effort never maps to a token budget. Unsupported or cross-family values are dropped with a warning (server log + OpenAI hipfire response metadata), not silently reinterpreted.

Chat framing and stop behavior are model-specific. Use the exact registry tag and do not assume Qwen <think> conventions apply to LFM, Gemma, Glimmer, or DeepSeek. The LFM framing smoke route is scripts/serve_harness.py with the exact lfm2.5:* tag; see VALIDATION.md.

Troubleshooting

The service has no authentication or TLS. Local chat should use a loopback bind; see SERVE.md before exposing it to a network.