Web UI

Quick start · Configuration reference

Open the browser interface to download models, chat, and change settings. The local service stays running in your terminal without loading a model:

snowllm webui                                  # foreground, opens the local Web UI
snowllm webui --no-browser                     # terminal/headless use

Running snowllm with no arguments shows the available entry points.

Manage models

From another terminal:

snowllm webui config --models-dir ~/models      # save the download/model root
snowllm webui pull qwen3.6-35b-a3b-fp8           # download in the background
snowllm webui logs --download                   # last 64 KiB of download output
snowllm status                                  # model and download status
snowllm load qwen3.6-35b-a3b-fp8 --max-model-len 32k
snowllm webui logs                             # last 64 KiB of inference output
snowllm unload

Browser interface

• Models: download recipes, load local checkpoints, save per-model load settings and manage storage. Downloads show progress and can resume after stopping.
• Chat: use browser-local sessions, template-based thinking controls, saved sampling settings, copy, regenerate and export.
• Language: English and Simplified Chinese follow the browser language; select manually in Settings.

Conversations and UI preferences stay in the current browser. The UI needs no external frontend service; model downloads need network access unless configured locally.

Deletion requires confirmation. Loaded models, their drafters and symlinks are protected; stop downloads before deleting files.

Monitoring

Chat replies show per-request prefill and decode throughput (tok/s), including thinking tokens. Prefill counts only tokens actually processed, excluding cached prefixes; decode excludes the first token emitted during prefill. Timing covers active engine steps, including scheduling and sampling, rather than browser/network latency. Statistics remain visible after refreshing the conversation.

The Overview shows engine queue/cache statistics and GPU utilization, device power, temperature, VRAM, and shared memory when Linux DRM/hwmon sensors provide them. Unsupported sensors show as unavailable. Device power follows the driver's sensor scope (on integrated GPUs it may include the APU). The inference listener exposes GET /metrics in Prometheus text format and GET /metrics?format=json for the same snapshot. The management listener proxies it at /daemon/metrics. Metrics reset when the model worker restarts; --stats-interval 0 disables aggregate engine accounting, but per-request chat timing and GPU sensors remain available.

Model lifecycle

Model changes are asynchronous: check snowllm status for draining, loading, ready, or failed (idle means no model is loaded). By default a change waits for active requests to finish before stopping the old model. snowllm load --force RECIPE or snowllm unload --force interrupts them. Put --force before the model name (snowllm load --force RECIPE). Only one model transition and one download can run at a time; downloads continue across model changes. Changing the model root affects future loads/downloads and does not move existing files.

The daemon runs two separate listeners:

Listener	Default address	Endpoints
Inference	127.0.0.1:8000	/v1/*, /health, /metrics
Management	127.0.0.1:8001	Web UI, /daemon/*

Clients keep using the same inference port and the existing /v1 APIs. The daemon streams requests to a separate inference process; while no model is ready, model inference endpoints return HTTP 503. The management service stays available if model loading fails. The default load timeout is 600 seconds (--load-timeout changes it), including worker startup. An active GPU wrapper can also consume this timeout.

The inference listener does not expose model management, filesystem browsing, or Web UI assets. Web UI chat uses dedicated same-origin endpoints on the management listener (/daemon/inference/models and /daemon/inference/chat/completions), which reuse the inference forwarding logic. Profiling controls are management endpoints: POST /daemon/profile/start and POST /daemon/profile/stop.

Configuration

Set model storage, the recipe catalogue and the kernel library in Settings. Changes apply to future loads and downloads; existing files are not moved.

See Configuration for JSON paths, overrides and CLI commands.

Service management

Choose the inference and management addresses independently:

snowllm webui --host 127.0.0.1 --port 9000 --admin-host 127.0.0.1 --admin-port 9001
snowllm status --admin-port 9001
snowllm load --admin-port 9001 RECIPE --max-model-len 32k
snowllm webui logs --admin-port 9001

For LAN inference, use --host 0.0.0.0; management remains on loopback. --admin-host supports 127.0.0.1, localhost, and ::1. Terminal management commands default to management port 8001; their --port option remains an alias for --admin-port. The daemon startup command's --port always means inference.

Starting a second daemon on the same management port reports the existing instance. Both ports must be available before either listener starts serving. Ctrl+C stops the service and its owned inference/download process trees; partial downloads can be resumed with another pull. The browser opens after both listeners are ready, or when connecting to an existing daemon, unless --no-browser is given. If opening the browser fails, the service continues and prints the URL. Management binds only to loopback and rejects cross-origin requests. Direct snowllm PATH|RECIPE serving remains available.