Models

This is where you set up what the agents run on: what your machine can handle, what you have downloaded, what is serving, which endpoints Minnow talks to, which model does which job, and what it all costs.

You come here to configure, then go back to Code and work. Open it from the app rail.

Section	What it is
Discover	Curated recommendations, Hugging Face search, file selection, and downloads
My models	Downloaded models and local loading
Storage	Model folders and Hugging Face credentials
Local Server	What is loaded, live load/inference chips, runtime log
Voice	Speech-to-text and text-to-speech models
Providers	Endpoints and encrypted API keys
CLIs	Claude Code, Codex, and Cursor subscriptions through installed CLIs
Routing	Which model handles which job
Routers	Shared capacity, sticky chat assignments, and model failover
Sampler	Temperature and sampling defaults
Thinking	Reasoning mode and budget
Usage & cost	Token totals and spend

Providers, Routing, Sampler, Thinking, and Usage & cost also appear under Models in the Settings sidebar. Routers has its own page in Models.

Local Server

This is the runtime dashboard: what is loaded, whether the process is healthy, and a live log. A loading card shows a modelled percent that actually moves (llama.cpp and mlx-lm do not print a weight-load percentage). Once the model is up, chips report prompt processing as a percent when the request came from Minnow, generated tokens as a count, and N queued when llama.cpp has more inference requests than free slots. mlx-lm has no server-side queue gauge, so that chip stays off.

Select a session to follow its runtime log and open its Overview. Expand Runtime snapshot to see the values that session started with: llama.cpp launch flags, MLX snapshot details, or readable MTPLX settings. The Connection area shows the selected ready model's base URL and model identifier, with copy controls and supported endpoints. Stop all stops the running models. Loading a model from Code does not yank you here; Local Server only comes to the front if you were already in Models.

Idle update_slots heartbeats are dropped from the log so they cannot drown the lines that matter.

Discover

Recommended is a short, hand-curated list for coding, everyday work, and reasoning. Each pick explains its purpose and shows a default Q4_K_M file size. Models within your estimated memory budget appear first. Filter by purpose or select Within memory budget to narrow the list.

Context tokens controls the memory estimate, starting at 16,384 tokens. Larger contexts need more memory. A model that needs a shorter context is labeled Context too long; Minnow does not silently shorten your selection to claim a fit. Estimates include runtime overhead and headroom. They distinguish GPU memory, unified memory, and slower CPU execution in RAM. Actual usage depends on runtime settings and other applications. Unknown metadata produces Fit unknown, not a promise that the model fits.

Hugging Face searches repositories by model name or owner/repository. Sort by downloads, likes, or recent updates, then use Load more models for the next page. GGUF is available on every platform; MLX is offered on Apple Silicon. Hugging Face results are not editorial recommendations.

Select Inspect files or Inspect to open the file inspector. Choose the exact GGUF quantization before downloading. Each choice shows its filename and download size; the memory estimate updates for that file. Split GGUF files appear as one choice with a total size and shard count. Missing shards disable that choice. Projector-only files are omitted. The model-card link opens the repository on Hugging Face. Access errors offer a route to Hugging Face settings; gated repositories may also require accepting the license on Hugging Face.

The Downloads shelf keeps progress, speed, and estimated time visible. Pause retains downloaded bytes; Resume continues the same job. Cancel & discard removes that transfer's artifacts. Failed transfers stay visible with an error and Retry download. Network failures retain partial files, while checksum failures discard corrupt bytes so retry can start cleanly. Completed jobs offer Open in My Models. Running jobs resume after restarting Minnow; deliberately paused jobs stay paused.

My models and Storage

Downloads from Discover appear in My models. Search or filter by maker, publisher, and format, or use Loaded and Needs attention to narrow the library. Quantizations of the same model share a row. Load uses that model's saved settings; the launch-settings button opens a large popover for adjusting them. Minnow registers the runtime as a provider so the model becomes available in Code.

Select a model to open its details panel. Overview shows runtime status, model size, trained context, capabilities, and connection copy controls when running. Expand Model & files for the repository, architecture, and local path, or Runtime snapshot for the current session's launch values. Inference provides per-model sampling controls with explanations and recommended presets. The panel's Load settings button opens the popover and returns to the same tab when you close it. The close button collapses the details panel.

Basic settings explains the everyday controls. For llama.cpp, choose context length, GPU layers, and KV cache precision, with a memory estimate above them. GPU Auto chooses a split for your hardware. For MTPLX, Basic exposes profile, context window, conversation-memory compression, and supported reasoning controls, with an estimate of unified memory. Available options and limits come from the model's MTPLX descriptor.

Advanced settings contains the detailed runtime controls: GPU selection, parallel slots, memory, speculative decoding, and extra arguments for llama.cpp; model, MTP heads, sampling, reasoning, performance, memory, persistent cache (SSD), thermal, and advanced controls for MTPLX. Switching between Basic and Advanced retains your edits. Changes save automatically and apply on the next load. Use defaults restores the selected engine's automatic or inherited settings. Close the popover with its close button, Escape, or a click outside it.

To remove a model from disk, choose the trash button on its row and confirm Delete model. If the row has a quantization picker, deletion applies to the selected quantization, including every shard of a split GGUF; other quantizations and shared image projectors are kept. MLX deletion removes the model folder and its contents, including all revisions of a Hugging Face cache repo. The confirmation shows the model and its current path. Eject a loaded model first, and cancel any unfinished download for its repo before deleting. Files in custom model folders are deleted from their original location too.

Parallel slots and KV cache. In the load popover, Unified KV cache makes the selected context one shared token pool. For example, 98,304 tokens with three slots allocates 98,304 tokens of KV capacity; concurrent requests share that capacity. With unified KV off, the context is per slot and the same selection allocates 294,912 tokens. The memory estimate updates when you change this setting. Existing saved context totals remain intact until you adjust the slider.

Loading GGUF on more than one GPU. The load popover's Advanced settings has a collapsed GPUs section. Check the cards that should run the model: the first you check is first in --device (so CUDA1 then CUDA0 means check CUDA1, then CUDA0). With two or more cards checked you can pick layer split (the default) or experimental tensor split, and drag per-card ratios. One GPU stays selected until you check another, so a second card stays free for the desktop. Loaded with lists Devices, Split, and Tensor split after a successful load. Extra llama-server args still override these fields.

Storage manages additional model folders and Hugging Face credentials. Model files are large; they are kept out of the small-backup path described in Where your data lives.

Runtime log. Local Server shows the selected runtime's complete recorded output and follows new messages. Minnow starts llama.cpp with verbose logging, including generated token IDs and text. Reload an already-running GGUF model to enable verbose output; an explicit verbosity setting in extra arguments takes precedence. Copy log copies the full received output, and Clear view clears the pane until more output arrives. Other engines show the messages their runtime emits.

MLX on Apple Silicon

On an Apple Silicon Mac, Minnow can also run MLX weights — Apple's Metal-native format. For the same quantization these are generally faster than GGUF on Metal, and the mlx-community and lmstudio-community accounts publish thousands of them.

MLX is Apple Silicon only. On Windows and Linux the option is not shown at all, and an MLX download is refused with an explanation rather than failing part way through.

Getting set up. The first MLX model you load asks to install the runtime. Minnow downloads a private Python environment and the mlx-lm packages — a few hundred megabytes, noticeably slower than the 20 MB llama.cpp install. The Python runtime is shared with Minnow's other managed servers, so it is only fetched once. You can also install it ahead of time from Settings → Servers → MLX.

Downloading. Search Hugging Face from Discover with the format set to MLX. An MLX model is a whole repository rather than a single file, so Minnow downloads the directory, skipping the original unquantized weights that many of these repos keep alongside the quantized ones.

Loading. MLX models appear in My Models with format MLX and a quant like mlx-4bit, and load the same way as GGUF — including a moving load percent while weights warm up. One difference is worth knowing: MLX runs as a single server that holds whichever model you asked for, so switching between two MLX models is a request rather than a process restart. The server keeps a model resident in memory after use; stop it from Settings → Servers when you want the RAM back. During a chat, prompt processing shows as a percent and generated tokens as a live count, same as GGUF.

Vision models are filtered out of MLX search. They need a different runtime that Minnow does not ship yet, and downloading 20 GB to hit a load error is not a useful way to find that out.

MTPLX on Apple Silicon

MTPLX runs MLX weights with multi-token prediction on Apple Silicon and macOS 14 or newer. Install MTPLX separately (pip install --upgrade mtplx); Minnow detects the installed executable and lists its model cache in My Models. Settings → Servers → MTPLX → Diagnostics shows the runtime's troubleshooting checks.

Select a model and open Load → Engine. Validated MTPLX rows offer Powered by MTPLX and MLX (no MTP acceleration). Missing files appear as a reason the model cannot load. In Discover, MLX snapshot downloads can target the MTPLX cache; the runtime validates them after downloading.

MTPLX controls use the same names as the MTPLX app's settings, so values carry over by name. Performance → Mode offers Fastest response, Handle multiple at once, and Long agent tasks; Concurrency cap sets how many completions run in parallel, and picking a Mode resets it, Decode batch max, and Admission window to that Mode's preset. The model supplies control bounds where available. Saved values outside those bounds are adjusted with a warning. Memory figures before loading are conservative estimates; the runtime determines its actual memory requirements.

If a matching MTPLX daemon already runs on the configured port or default port 8000, Minnow connects to it and marks it External daemon. Its existing settings remain in effect. Eject disconnects Minnow without terminating that daemon. Otherwise Minnow starts a process, manages its idle timeout, and stops it on eject or app shutdown. Requests sent directly to its endpoint also renew the idle window. A daemon that crashes after a healthy run receives one automatic restart with the same port and settings. Local Server shows decode/prefill speed, time to first token, cached tokens, MTP depth and active requests.

Powered by MTPLX.

Providers

A provider connects Minnow to a model service. You can have as many as you like, enabled independently.

• Local runtimes — LM Studio on http://localhost:1234 and Ollama on http://localhost:11434 are detected automatically when they are already running on their default ports.
• Cloud APIs — one-click presets for OpenCode Go/Zen, Anthropic, DeepSeek, GitHub Copilot, OpenRouter, OpenAI, Groq and Mistral, plus a custom option.
• Managed — anything you serve from the Library.

API keys are encrypted at rest with AES-256-GCM. Losing the key file in your Minnow home means re-entering them.

Select a provider row to open its connection card. Changes save automatically when a field loses focus or you press Enter. The status below the fields confirms when they are saved. Advanced settings contains API paths, authentication headers, gateway routing, pricing, and capability checks. CLI connections have a Manage CLI link.

Choose Add provider for a local or cloud preset, or Custom endpoint for another server. Provider IDs are filled in automatically and can be edited under Advanced settings.

Test connection checks the saved endpoint and reports the available model count or a connection error. Wait for edits to finish saving before testing. This check does not generate a model response.

Full walkthrough: Connect a model.

CLIs

Use an installed Claude Code, Codex, or Cursor CLI as a model provider. Open Models → CLIs, scan for installations, and expand a CLI row to enable it. Its models appear in the normal model picker and can be assigned to chats and work agents.

If a CLI is missing, choose Install. Minnow opens Terminal and runs the vendor installer for that shell — including Cursor's PowerShell installer on Windows. When the installer finishes, return and choose Scan again. Update runs the CLI's own updater (claude update, codex update, or cursor-agent update) in Terminal, so you get each vendor's latest release as soon as it ships. Minnow first closes idle CLI processes it started so the update can replace their files; chats that are running keep going, and closed conversations resume on the next message. When the update finishes, choose Scan again to pick up the new version and any models it adds. Sign in shows connection progress here while you approve access in your browser. Use Open sign-in page if the browser does not open. Minnow verifies the login automatically; you can cancel or retry without leaving Models. If Claude shows a code after browser approval, expand Browser gave you a sign-in code? and paste it here. Verify checks a login you completed elsewhere. Scanning and verification do not generate a model response or consume an inference request. Some credential stores cannot report login status; Sign-in unverified means Minnow could not confirm it. Credentials stay with the CLI. Codex sign-in uses its native auth.json file store so isolated requests can reuse the login; a keyring-only login needs this sign-in step once. Your saved CLI configuration is not changed.

Each expanded row shows connection controls, concurrent runs, and the context window. Claude Code retains a dollar-budget field for the legacy headless connection; leave it empty for interactive chat. Advanced settings contains the executable path override and Allow helper tasks. Changes save automatically as you edit; the status below the fields confirms when they are saved. If a save fails, your input stays in place and Retry sends it again.

Concurrency defaults to one; further requests wait in order, and Stop also cancels a queued request. Background use is off by default so title generation, editor completions, and similar utility jobs do not silently use your subscription. Assign those jobs another provider or explicitly enable helper tasks. Claude Code's interactive connection does not support the CLI dollar-budget flag. If a budget is configured, Minnow asks you to clear it before starting interactive chat; it never silently ignores the cap.

Context window (tokens) defaults to automatic model discovery. Enter 1,000–1,000,000 tokens to set an explicit budget, or clear the field to restore automatic. Codex receives this window in its native configuration; select a value your model supports. Claude requests extended context above 200,000 for Sonnet and Opus when your account supports it; Haiku remains capped at 200,000. A value at or below 200,000 holds Claude to its smaller native window. Cursor's setting only lowers Minnow's budget below the model's advertised capacity; it cannot enlarge Cursor's native window. Changes apply when the next generation starts.

Conversations continue automatically while each generation request ends when the CLI finishes its response. Claude Code runs an interactive session behind normal chat and requires CLI 2.1.289 or newer; Codex uses a managed app-server connection and requires CLI 0.153.4 or newer. Ordinary follow-ups send only new input. Minnow retains up to eight idle processes per provider for five minutes. Clean, matching saved conversations resume after eviction or a Minnow restart. Cursor uses its ACP session protocol when its installed version passes the session-loading handshake and Minnow's tool-isolation requirements; otherwise Minnow selects isolated replay before sending a prompt and shows the reason in the CLI view.

Claude sends text and images through its private connection, with background prompt suggestions disabled. Minnow continues to handle tool execution and approvals. Token counts come from completed native responses; missing dollar-cost or quota measurements remain unavailable.

The first interactive connection keeps Claude's selected display theme and continues its informational security notice. Startup progress and errors appear in the CLI view. If Claude needs sign-in, use Sign in here. If it needs additional first-run setup, run claude once in Minnow Terminal, finish those prompts, then retry your chat.

Claude also saves checkpoints at tool handoffs. After an interruption, a matching conversation resumes from the checkpoint when all pending tool results are recorded, sending those results without replaying the earlier conversation. If the saved history or configuration no longer matches, Minnow rebuilds from its recorded history. CLI providers receive their permitted tools together to keep the conversation stable as work progresses.

Edits, regeneration, context trimming, instruction or tool changes, workspace changes, account changes, and interrupted or modified native histories reconstruct the conversation from Minnow's recorded history. Stable instructions stay separate from current-turn context; changing an opaque custom system prompt also reconstructs the conversation. Minnow returns real tool results to the same native turn when available and reconstructs with recorded results after an interruption, without executing those tools again. Pending tool requests use their own timeout. The CLI can request up to eight independent Minnow tools in one batch; Minnow applies the usual mode restrictions, approvals, tool cards, user questions, and board reporting. Stop cancels the generation and closes its process. Closing a workspace releases its processes while retaining valid saved conversations; deleting a chat removes its saved binding. Unused CLI session caches expire after 30 days and are excluded from backups.

When a chat uses a CLI provider, the CLI button at the upper right of the conversation shows the current process output as it runs. Select Chat to return to the conversation. This view appends native JSON output as it arrives and recovers a recent snapshot after reconnecting. Claude and Cursor also show recent stderr. The view is read-only, and messages are still sent through the chat composer. The recent output remains available until Minnow restarts or the capture is replaced by a new process.

The CLI view distinguishes Ready for next message, Waiting for Minnow tool results, Resumed saved conversation, and Conversation rebuilt, with the rebuild or fallback reason. Usage covers the latest generation and separates uncached input, cache reads, cache writes, and output where the CLI reports them. Missing counts or cost show as unavailable. When Claude reports one cost for several requests across tool steps, it is labeled Native turn total; an unavailable per-generation cost is not replaced with that whole-turn amount. Token totals and reported dollar cost are not subscription quota measurements. Conversation persistence enables cache reuse but does not guarantee reduced subscription usage or preserve a provider cache indefinitely.

Claude Code offers moving aliases and pinned version choices supported by the detected CLI version. Alias resolution can vary by provider; a pinned model ID keeps the version fixed. Claude Code supports image attachments and reasoning effort. Codex supports reasoning effort, image attachments, and tool screenshots when the selected model accepts images. It lists models and their image support through the installed CLI, independently of the Codex desktop app. Model discovery does not generate a reply; choices are cached for up to five minutes. If a saved model is no longer supported, refresh the model list and choose a supported model, or update the CLI using Install; Cursor uses its CLI's model defaults, including the models your account lists. Unsupported sampling options are not forwarded. CLI conversations travel over standard input with an 8 MB transcript bound. Codex submits only new input while its accepted conversation remains available. CLI access, available models, and account limits follow the installed CLI and your account.

Minnow disables Claude Code's automatic session-title inference. Chat titles use Minnow's configured title provider and background-use controls.

Account usage

Codex, Claude and Cursor connections show subscription allowance in Models → CLIs. Open a CLI row for each quota window, its remaining percentage, reset time and last refresh. Cursor reports its monthly included usage as Total, Auto and API meters that reset at the end of your billing cycle. When a Code chat uses one of these CLIs, the Usage button beside the model opens the same account details. After loading, the button shows the remaining percentage for the most-used window.

Usage covers your account, including activity outside Minnow. Checking it does not generate a model response. Visible usage indicators refresh about once a minute; Refresh checks again subject to the provider's retry interval. If a refresh fails, Minnow labels recent measurements as last-known usage. An asterisk beside the composer's percentage also marks last-known data. Missing limits show as unavailable rather than zero usage.

Codex requires a ChatGPT login; Claude requires a current subscription OAuth login. API keys and other API configurations do not provide subscription quota. Login credentials stay on the tool server. Claude uses the usage endpoint used by its CLI; changes to that endpoint can temporarily make usage unavailable. If your login expires, verify or sign in through the CLI and refresh usage.

Routing

Open Models → Routing. Models by task shows the active conversation and dedicated models for utility tasks, UI design, goal evaluation, and editor assistance. Each row shows whether it uses a dedicated model or follows the chat or task. Choose a model in the shared searchable dropdown, or use Follow chat model / Follow each task’s model to clear an override. Saved models remain visible when a provider is offline. Changes save automatically.

Options holds task-specific settings and fallback models. The Fallbacks tab controls shared backup models, whether fallback is enabled, and cooldown after a host fails. Use the arrows to set backup order and Remove to remove an entry. The add control disables at the chain limit. Switching tabs preserves the controls and edits already on the page.

Instead of one model doing everything, bind models to roles: main chat, utility tasks, the /goal evaluator, the UI Designer runtime, and each work agent and sub-agent type such as builder, planner, reviewer, and researcher.

Two bindings are worth setting deliberately:

• Utility tasks. One shared model handles chat titles, prompt and issue expansion, and git commit messages. Leave it unset to keep using the current composer, top-bar, or editor model for each task. A small fast model is often enough.
• Goal evaluator. This one judges whether your /goal condition is genuinely met. A weak evaluator rubber-stamps broken work, which is worse than no goal at all.

A common arrangement is a fast local model for routine turns and a capable cloud model bound to review, research and evaluation.

Model pools

Open Models → Model pools and choose New model pool. Name the pool, then use Add models to browse and search models from My Models and your connected providers. Choose Add beside each model you want; models already in the pool show Added. Choose Done to return to the lineup. Local llama.cpp and MLX runtime catalogs are represented by their weights in My Models, just as in the chat picker.

The Model lineup stays visible. Drag a model's handle to change its rank, use the arrow buttons, or press Alt+↑ / Alt+↓ while a row has focus. Choose the model dropdown to replace an entry; it uses the shared searchable model picker. A provider/model pair can appear only once in a pool. Active / Paused controls whether a model receives work. The − / + buttons adjust concurrent generation slots; More offers an exact limit (1–100) and Remove from pool. Removing an entry keeps the model in your library. Changes save automatically, with Retry save available if saving fails.

When a chat is assigned a My Models entry that is not loaded, Minnow loads it before generating. If another local model is still producing a response, the router waits for that work to finish, then unloads it if residency requires and loads the assigned weights. Idle TTL (twenty minutes) still applies. Cloud and LM Studio entries are unchanged.

Prefer first prefers the first eligible model with free capacity. Share work assigns new chats using rank weights: a three-entry router uses weights 3, 2, and 1. Once assigned, a chat keeps that model. If its model is busy, the chat waits in a FIFO queue even when another entry has capacity. If all entries are busy, new chats queue too. Streaming and non-streaming generations each occupy one slot.

Match the task chooses a model and reasoning level once for each chat. Choose a model in the shared Task evaluator dropdown. For an unlisted ID, expand Use a model ID not listed here, select its provider, enter its ID, and choose Use evaluator. A decision model such as TypeSafe Jev (typesafe/jev-1.13 on OpenRouter) or Cloudflare Clef (on Workers AI or OpenRouter) answers through the Decisions API with real confidence scores, and any other chat model can judge too by answering as JSON — pick something small and fast. Each entry gets a Tier (Fast, Balanced, or Frontier); Auto guesses it from the model name. On your first message, the evaluator reads your request and chooses a tier and a reasoning level. The chat keeps that choice for follow-ups and tool rounds, including after a reload or restart. Open the composer model menu and choose Re-evaluate to request a fresh choice on your next message. The router runs the highest-ranked eligible entry in that tier, or the nearest tier when none is eligible, preferring a stronger one, and sets the reasoning level to the closest one the model offers. The model stays assigned while eligible; an unavailable or failed model can still move to another entry. When a re-evaluation is unsure (below 55% confidence) the chat keeps its current tier and reasoning. The evaluator has up to four seconds for a decision model, 30 seconds for a chat model, or 60 seconds for a CLI model, including CLI startup. CLI evaluators require background use to be enabled in Models → CLIs. If the evaluator is unset, fails, or exceeds its time limit, the router keeps the last choice or uses rank order. That fallback stays in place until you choose Re-evaluate. A pinned override skips the evaluator.

While the evaluator runs, the composer's model label shows Evaluating. Its choice appears immediately, for example High reasoning · deepseek-v4. During re-evaluation, the current model and reasoning remain visible until the new choice arrives. If the provider does not advertise reasoning levels, the label shows the requested effort, for example High requested, and the provider uses its default. Models known to lack reasoning show No reasoning. Hover the label for the evaluator, confidence, tier, requested effort, and explanation. The live view shows each chat's last judgment.

Pools appear in the normal model picker with a Model pool label. Use for new chats sets the workspace's default pool; existing chat bindings stay unchanged. A chat's picker shows the router and its current provider/model assignment after a request starts.

The live view shows active and queued chats, their target models, and slot usage. Inspect a model card for average generation latency, error rate, reported tokens, and estimated cost when provider pricing and token counts are available. The chat's override selector pins an entry until you choose Model pool assignment to clear it. An override can fail over if its model fails; the replacement becomes sticky while the override remains marked until cleared.

Provider errors and unavailable models trigger failover. A response interrupted partway through restarts on another eligible entry with a visible warning; failed text, reasoning, and incomplete tool calls are discarded. Each request attempts a provider/model pair at most once. Stop cancels the request without failover. If no eligible entries remain, Minnow asks you to check the router's entries and provider configuration.

Pool configurations, defaults, and chat assignments are saved per workspace. Activity and telemetry last for the current server session. On narrow screens the activity and model views stack; reduced-motion settings replace moving connections with static indicators.

Pool active / Pool paused enables or pauses the whole pool. Pausing the default pool clears that default. Delete pool asks you to confirm before removing the pool; it keeps the models in your library.

Sampler

Temperature, top-p, top-k, min-p, repeat penalty, presence penalty, max tokens.

For a downloaded model, open its inspector and choose Inference → Recommended preset. Select a preset and click Apply preset to fill and save the recommended sampler values for that model. Every field stays editable, and your later changes are saved. Selecting a model or choosing a preset in the dropdown leaves your current values in place until you apply it.

Presets cover Qwen3, Qwen3.5, Qwen3.6, Qwen3.8 (including Flash Next), Qwen3 Coder Next, Gemma 4 instruction models, and DeepSeek R1 (including distills). The Model guidance link opens the official recommendation. Thinking and non-thinking presets describe the sampling to use with that mode; set the thinking mode separately in the composer or Thinking settings. Applying a preset preserves your output limit and any sampler fields the recommendation does not specify. Empty fields inherit global defaults. Models without a known recommendation keep the same editable fields.

For example, Thinking / precise coding for Qwen3.6 sets temperature 0.6, top-p 0.95, top-k 20, min-p 0, presence penalty 0, and repeat penalty 1.0. Explicit zero and neutral values override inherited sampling values on supported runtimes. Providers may support only some sampler parameters.

Thinking

Reasoning mode and token budget for models that expose reasoning. Minnow displays reasoning separately from the answer and times it — the "Thinking…" clock covers reasoning only, stopping when tool calls begin, so the number means something.

For models that expose named reasoning levels, open the model picker and set Reasoning default below the model list. The choice is saved for that provider and model, then applied when you select it in a chat. Choose Model default to use the level advertised by the provider again. The composer control can still override the level for the current chat.

Usage & cost

Token totals for the active chat and for the workspace session. Enter per-million pricing for your models and it becomes actual spend rather than an abstract count.

Voice

Built-in dictation automatically downloads a compact speech model on first use. System voices provide read-aloud without a model download. Optional larger Whisper and Qwen3-TTS models use the advanced local Python runtime; external voice providers are also supported.

See Voice.

Choosing the model for a turn

Control	Scope
Menubar model chip	Global default — what new chats start with
Composer picker, Ctrl+M / Cmd+M	This chat only

Local runtimes expose Load and Unload in the composer picker, acting on the model that chat is bound to. The tray menu can unload local models without opening the window — useful when you want your VRAM back. Live load and inference numbers live on Local Server.

When the picker is empty

1. Is the provider process running, with a model loaded?
2. Is the base URL right, including /v1 where required?
3. Open the provider card and choose Test connection.

[providers] fetch failed at startup is normal when a local runtime is not up yet.

Related

• Connect a model
• Voice
• Settings app
• Troubleshooting