Chat

hipfire provides two native Rust chat surfaces:

• hipfire tui (or bare hipfire) opens the full ratatui application and its Chat tab. This requires the hipfire-tui binary; when it is not installed the command exits with an error that names the build command.
• hipfire chat [model] is a lightweight line-oriented multi-turn client for terminals and scripts. It is part of the hipfire binary and does not need hipfire-tui.

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.

Chat-tab slash commands run locally and never send a message (a message whose text must begin with / is sent as //text, which transmits /text):

The transcript follows new output while you are at the bottom: Up/Down and PgUp/PgDn scroll (also while a reply streams), Home jumps to the top and End resumes following. Typing keeps working during a reply; only Enter waits for it to finish, and Esc stops it (the reply is marked [stopped]). Reasoning text is shown dimmed under a thinking label. The token count under a reply is the serve’s usage.completion_tokens when the serve reports it, otherwise the number of streamed chunks, labelled as chunks.

The TUI checks the serve’s /health every 2 s on every tab. When it is not online, Enter queues your prompt (Esc cancels it and puts the text back), starts hipfire serve -d only when the first check says the serve is down and no serve process exists, and sends the prompt by itself once the serve is online. 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 (drops every message except the system message), /exit, and /quit. Flags:

[model] is optional and defaults to the configured serve.default_model.

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.