AI Models
Open in CMSThe AI Models page decides which AI models the organization can use and which model powers each platform feature by default. Owners and Admins also connect the organization's own provider keys and their Claude or ChatGPT subscription here. For step-by-step guides, see Bring Your Own Key and Use Your Claude or ChatGPT Subscription. Open it with the AI Models tile on the organization's management menu.
Access
| Owner/Admin | Can view and change everything on this page. |
| Other Roles | The AI Models tile is not shown. Opening the page by its address shows "Only organization Owners and Admins can open AI Models." |
Tabs
The page has three tabs. The open tab is kept in the address (?tab=models, ?tab=sources, ?tab=tasks), so a refresh or the browser's Back button returns to the same tab.
| Models | Every model the organization can use, with its on/off switch. Opens by default. |
| Sources | The organization's Subscriptions and provider API keys. Shown for every organization unless Qlar has turned both API keys and subscriptions off for it (while subscriptions are turned off, it still shows for an organization that has a connected computer); otherwise ?tab=sources opens Models. |
| Tasks | Which model does each AI job: agent features, chat helpers and builder assistants. |
| On a phone | The tabs share the screen width evenly. |
Offline Computer Banner
Above the tabs, one warning per connected computer that is offline while its models are switched on: "Office PC is offline, so sonnet-office can't answer right now." It adds "It is used for …" when one of those models is chosen for a feature or task. Bring it back online opens the computer's Bring it back online panel on Sources.
Models Tab
Models are grouped by where they come from. Each row shows the model's alias, a one-line description, chips for what it accepts and produces (for example Image in, Document in, Audio out, or Text only), and a switch.
| Filter chips | All, Qlar, API key, and Subscription, each with its model count. API key is hidden where Qlar has turned API keys off for the organization, Subscription where it has turned subscriptions off. |
| Search models | Narrows the rows by alias or description. |
| Qlar | Built-in models, billed with Qlar credits. The description includes the credit multiplier, for example "3× credits". |
| Your API keys | Models reached with the organization's own provider keys. No Qlar credit is charged for these. Add API key opens the Add API Key dialog. |
| Your subscriptions | Models from your own Claude or ChatGPT subscription, on the organization's connected computers. No Qlar credit is charged for these. Connect a subscription opens the Connect a Subscription wizard. Each row shows the tool and model, then where it runs: ● Office PC while online, ○ Office PC, offline 3 h ago while not, or Office PC · not offered by this computer when the computer no longer offers that model. |
| Model settings | The slider icon on a row of Your API keys or Your subscriptions opens the Edit Model dialog, for example to set its reasoning effort. A model with an effort shows it in its description, for example "Effort: Low", and a model with Fast mode on shows "Fast mode". On a phone, Model settings is in the row's sheet. |
| Switch | Turns the model on or off for the organization. A model turned off is hidden from every agent. |
| Default chip | Default followed by the features this model powers, for example "Default · Conversation". |
| Locked Switch | A Qlar model chosen for a feature or task on Tasks cannot be turned off. Choose another model there first; the tooltip names the features and tasks. |
| On a phone | Rows show only the name, description and switch. Tap a row to open a sheet with its capability and default chips; Close dismisses it. |
Sources Tab
The tab has three sections, Subscriptions, API keys and Service keys. The first two are shown only where the organization may use them; Subscriptions also stays visible while subscriptions are turned off for an organization that still has a connected computer.
Subscriptions
| Connect a subscription | Opens the Connect a Subscription wizard in place of the list. |
| Setup rule | Under the heading: "Subscriptions can be connected only in an Individual organization, on up to 2 computers per organization." |
| Can't connect | In a Collective organization, or once the organization has 2 connected computers, Connect a subscription is greyed out and a note says why: "Subscriptions can be connected only in an Individual organization. Switch to your Individual organization to connect one." or "This organization already has 2 connected computers, the most allowed. Remove one before connecting another." The same applies to Connect a subscription on the Models tab. Computers an organization has already connected keep working and can be removed. |
| Turned off | "Subscriptions are turned off for this organization, so their models can't be used. Everything comes back as it was when Qlar turns them on again. You can still remove a connected computer." Cards are read-only and offer Remove only. |
| Empty State | No subscriptions yet. Connect a computer to use your own Claude or ChatGPT subscription instead of Qlar credits. |
| Card | One card per connected computer: its name, a status chip, ● Online or ○ Offline · last seen …, then the computer's name, system, Qlar connector version and "n of 2 jobs running", a chip per tool with how it signs in (for example Claude Code · subscription (Max)), and its models. |
| ⋯ menu | Rename, Move to another computer (approved computers), and Remove…. |
Card status and the button it offers:
| Status chip | Meaning | Button |
|---|---|---|
| Connected · ● Online | Approved and running. | Models · n: add or remove its models. |
| Connected · ○ Offline | Approved, but not running or not reachable. | Bring it back online |
| Waiting for the command to run | The setup command hasn't reported in. Its code can't be shown again. | Get a new code |
| Code expired | The command was never run within 15 minutes. | Get a new code |
| Waiting for your approval | The computer enrolled; its fingerprint waits for approval. | Review fingerprint |
| Moving to a new computer | A move is waiting for its command on the new computer. The current computer keeps working. | Get a new code |
| New computer waiting for approval | The new computer enrolled; the current one keeps working until you approve it. | Review fingerprint |
| Move: code expired | The code for the new computer expired before it was used. | Get a new code |
| Revoked | The computer can't connect any more. | Remove |
| No longer registered | Qlar's subscription service no longer knows this computer. | Remove |
| Status unavailable | Qlar couldn't reach its subscription service to check this computer. | Retry |
Bring it back online
Opened from an offline card or the offline banner.
| Same computer | The commands to start the Qlar connector again: qlar-inference-gateway run in a terminal, or the restart command for a service (Windows schtasks /Run /TN "Qlar Inference Gateway", macOS launchctl kickstart -k gui/$(id -u)/ai.qlar.inference-gateway, Linux sudo systemctl restart qlar-inference-gateway), and the install-service command so it survives the next restart. The card turns green as soon as it reports in. |
| New or reset computer | Get a new code starts a move to another computer. Its models, the agents using them and their settings stay as they are. |
API keys
| Key card | One card per key: its name (or the provider when it has none), then provider, endpoint (the Server URL of an OpenAI-compatible server), masked key (no key for a server added without one) and how its models are called (Responses API or Chat Completions, set by Qlar). Below, one row per model: the alias, the model name and description, capability chips, Not checked by Qlar for a model typed in, Retires on … / Retired since … when the provider announced a shutdown date, and Turned off when the model is switched off under Models. |
| Add models | Opens the Add API Key dialog for that key, models only. Models already added show Added as alias, ticked and fixed. |
| Edit key | Opens the Edit Key dialog. |
| Edit (model row) | Opens the Edit Model dialog. |
| Add API key | Opens the Add API Key dialog. |
| Older entries | A key added before a key could hold several models shows as one card per model (alias, provider, model, endpoint, masked key) with Edit, until Qlar groups it. |
| Empty State | No API keys yet. |
Service keys
Keys of the organization's own accounts for services that are not AI models. Today: Exa, for web search. See Web search on your own Exa account.
| Key row | Its name, Exa · web search, the masked key (last 4 characters), and Used for web search while web search runs on it. |
| Add Exa key | Opens Add Exa key: Name (optional) and Exa API key. Check and save runs one small search with the key first; a refused key isn't saved ("Exa refused this key…"). Adding a key doesn't move web search: choose it on the Web search row of the Tasks tab. |
| Delete | Asks first. Deleting the key in use moves web search back to Qlar's Exa account, or turns it off while only your own models are allowed. The key is removed from Qlar; your Exa account is not changed. |
| Empty State | No service keys yet. |
Connect a Subscription Wizard
Opens in place of the Sources list, titled Connect a subscription (Move name to another computer when moving, Models on name when adding models), with "Step n of 4" and a progress bar. × closes it; setup continues from the computer's card. A refresh keeps the wizard open (?gateway= in the address).
| Step | Fields and buttons |
|---|---|
| 1 Where | A note at the top repeats the rule: subscriptions only in an Individual organization, on up to 2 computers per organization. Name (required, at most 100 characters, shown to your team in model lists) and Where will it run?: Windows (Windows 10 or 11, x64), macOS (Apple silicon), Linux (x64 or arm64). Create code adds the computer and creates a one-time code that expires after 15 minutes. |
| 2 Run | The command for the chosen system, with links to switch to the other systems. Copy command; on a phone or where sharing is supported, Send to my computer…. The One-time code with its own copy button, and the shorter enroll command for a computer that already has the Qlar connector. A countdown while waiting for the computer to report in. Get a new code when the code expired or the page was reloaded. Cancel deletes a computer that was never approved. |
| 3 Confirm | A computer has enrolled — confirm it is yours: the computer's name, system and Qlar connector version, and the Key fingerprint (SHA-256) as 4 rows of 4 groups. Fingerprint matches — approve, or It doesn't match — cancel. If the key changes while you compare, Show the new fingerprint. |
| 4 Sign in | How each tool is signed in (subscription, API key, not signed in with the command to run, or Waiting for sign-in…), then Choose models to make available: a checkbox and a name for each model. Names use lower-case letters, digits, ., _ or -, no +, and must not clash with another model. A model of a tool that isn't signed in shows "sign in to use it". Keep it running after a restart shows the install-service command for the system. Finish adds the ticked models; Skip for now closes without adding. |
Rename and Remove Computer Dialogs
| Rename name | Name (at most 100 characters). Save. Only the name in Qlar changes. |
| Remove name? | Explains that the computer is disconnected and can't be used again, and lists its models. Then what else changes: the features whose default goes back to the Qlar default (which uses Qlar credits), the agents whose models fall back to the organization default, and the agents whose voice calls stop until another voice is chosen (they don't move to a Qlar voice). Keep it or Remove computer. A computer that was never approved is removed without this dialog. |
Add API Key Dialog
Titled Add API key (Add models to name from a key's Add models, which shows only Models). One key, any number of models.
| Provider | OpenAI, Azure OpenAI, Azure AI Inference, DeepSeek, Anthropic (Claude), Google Gemini, OpenRouter (many vendors, one key), or OpenAI-compatible server (Kimi, Groq, vLLM…). Changing it clears the ticked models. On OpenRouter, a note lists what it doesn't support (realtime voice, Fast mode, embeddings) with a link to OpenRouter keys, and Models shows the OpenRouter list instead of Qlar's: Load models, then Search models, a kind filter (Chat, Text to speech, Speech to text, Image, All), up to 50 models with their price, badges (Image input, PDF input, Tools, Effort, voices) and warnings ("No tool calling", free variants), and Add as … per model. |
| Name (optional) | How the key is shown, for example Marketing OpenAI. An OpenAI-compatible server without a name is shown by its host. |
| Endpoint | Azure providers only, required. |
| Response API endpoint | Azure AI Inference only, required. |
| Server URL | OpenAI-compatible server only, required: the base URL before /chat/completions, for example https://api.moonshot.ai/v1. Must be a public https:// address; http://, local, private and Qlar addresses are refused, in the dialog and again by Qlar. |
| API key | Required, except API key (optional) for an OpenAI-compatible server ("Leave empty only if your server needs no key"). Stored encrypted; afterwards only a masked part is shown. |
| How models are called | "Qlar calls these models with the Responses API, set automatically" (the Chat Completions API for DeepSeek, Azure AI Inference, Anthropic, Gemini, OpenRouter and OpenAI-compatible servers), and that usage through your own key uses no Qlar credit. There is no invocation method or credit ratio to fill in. |
| OpenAI-compatible note | "Qlar doesn't check this server…": it must answer the OpenAI Chat Completions API and support tool calling if your agents use tools; reasoning models spend part of the reply limit on thinking. |
| Load models from the server | OpenAI-compatible server only, in place of Qlar's list ("Qlar keeps no model list for your own server"). Asks the server for its models and shows each as a chip; + model adds a Your own model card with that name, and an added model shows ✓. A failure is shown in red, for example "Use a public address, not a private one" or "The server answered 401: … Check the API key." |
| Models | Qlar's list of known models for the provider: name, model name, description, capability chips (an embedding model shows "Embeddings · n dimensions" instead) and Retires on … when the provider announced a shutdown date. Models the provider no longer offers are not listed. "n selected" counts what will be added. |
| Alias (optional) | Shown for a ticked model. Leave empty to use the model name. Letters, digits, ., _, : or -, starting with a letter or digit; no +. A name already used by a Qlar model, another model or another row shows "Already used by another model. Enter an alias." and blocks saving. |
| Deployment name (optional) | Azure OpenAI only; defaults to the model name. |
| Reasoning effort | Shown for a ticked reasoning model from Qlar's list: Model default (default), then only the efforts that model accepts. Not shown for a model without reasoning effort (for example GPT-4.1). |
| Reasoning effort (optional) | On a Your own model chat card: free text as the provider documents it, for example low; empty is the model's default. "DeepSeek takes none, high or max" on a DeepSeek key, "Claude takes low, medium, high, xhigh or max" on an Anthropic key, "Gemini takes minimal, low, medium or high" on a Gemini key, and on an OpenRouter key the efforts OpenRouter lists for that model ("This model takes: …"). A value with spaces or other characters blocks saving. |
| Fast mode | A switch under a ticked OpenAI model that supports Fast mode, and on a Your own model chat card on an OpenAI key. Off by default. |
| Add a model that isn't listed | Adds a Your own model card: Model name (required, exactly as the provider names it), Alias (optional), Deployment name (Azure OpenAI), Accepts (Image, Audio, Video, Document) and Produces (Image, Audio, Video) switches, and Document types when Document is on (at least one of PDF, DOCX, CSV, TXT, XLSX, PPTX). For an OpenAI-compatible server only Accepts (besides text): Image. For OpenAI, Azure OpenAI and an OpenAI-compatible server, Embedding model (for knowledge search) turns the card into an embedding model: the capability switches give way to Dimensions (required) and Detect, which embeds one short text with the model to read its vector size. × removes the card. On providers that serve them, Speech-to-text model (for dictation and voice notes), Text-to-speech model (for spoken replies and voice previews) and Realtime voice model (for voice calls) (OpenAI only) make the card a model of that kind, offered only for its Voice task. |
| Warning | With a model of your own: "Qlar doesn't check models you add yourself: make sure the model name and what it can accept and produce match the provider's documentation. A wrong name only shows up as a failed reply." |
| Add key / Add models | Saves the key and the models in one step. Disabled until at least one model is chosen and every alias is valid. |
Edit Key Dialog
Titled Edit name. The provider can't be changed.
| Name, Endpoint, Response API endpoint, Server URL | As when adding. A changed Server URL applies to every model of the key. |
| API key (leave blank to keep current) | Shows "Current key ••••last 4" (or "No key saved: Qlar calls the server without one"). A new key applies to every model of the key, listed under "Used by …". |
| Remove the saved key (the server needs none) | OpenAI-compatible server with a saved key only. Drops the key from the server and all its models on Save. |
| Test connection | OpenAI-compatible server only. Checks the URL and key (the saved one when the field is blank): "Connected. The server offers n models", or the reason it failed. |
| Remove key | Asks "Remove this key and n models?"; Yes, remove removes the key and its models. Refused while one of them is a feature default. |
| Save | Saves the key. |
Edit Model Dialog
Titled Edit "alias". For a model of a key it shows the provider, masked key and how the model is called, and only the model's own fields:
| Native model name | Fixed for a model from Qlar's list; editable for a model you typed in. |
| Deployment name | Azure OpenAI only. |
| Short description | Optional summary shown to users. |
| Reasoning effort | Chat models only. A model from Qlar's list, or one on a DeepSeek, Anthropic or Gemini key: a list with Model default (default) first and the efforts the model accepts; a model without reasoning effort shows "This model has no reasoning effort setting." A model you typed in: free text, empty for the model's default. Applies to every call on the model. |
| Fast mode | Chat models that have it: an OpenAI model from Qlar's list that supports it, an OpenAI model you typed in, or a subscription model the computer reports with Fast mode. The help text says who pays. Applies to every call on the model. |
| Capabilities | Fixed for a model from Qlar's list ("from Qlar's list of known models"); for a model you typed in, the Input and Output switches and Document types, with "Qlar doesn't check models you add yourself". On an OpenAI-compatible server only Input › Image can change. |
| Delete | Asks "Delete this model?"; Yes, delete removes this model only. |
| Save | Saves the model. |
A subscription model opens a short form: the alias, "Subscription · computer · model", and Reasoning effort with the efforts the computer reports for that model (the efforts a subscription can send while the computer is offline). Only the effort and, for a model the computer reports with Fast mode, Fast mode can change there.
An older entry (one card per model) shows the earlier form instead: Provider, Invocation method, Endpoint and API key are editable there, and there is no credit ratio.
Tasks Tab
Choose which model does each AI job. ?tab=defaults, the address of the earlier Defaults tab, opens this tab. Each change saves immediately, and a toast confirms it (for example File check now uses "my-gpt".).
Where your data goes at the top counts how many AI tasks run on models you control ("3 of 25 AI tasks run on models you control"), with a bar split into Qlar platform, Your API key and Your subscription.
AI calls in the last n days counts the organization's AI calls on its own models and on Qlar's, by group, for 7 days, 30 days or 90 days ("1,240 calls on your models · 3 on Qlar's", then "n yours · n Qlar" per group). See details lists each job and model with its count. Under Only use our own models it lists the calls that still ran on Qlar's models on or after the day the rule was turned on. "The call counts can't be read right now." when the counts are unavailable.
Under it, the Owner's Only use our own models switch (Only Use Your Own Models):
| Switch | Off: "When on, every AI job must run on a model of your own. Nothing falls back to Qlar's models." On: "On: every AI job runs on a model of your own. A job without one stops instead of using Qlar's models." Below it, who turned it on or off and when. |
| Admins | See the switch greyed out with "Only the owner can change this." |
| Only use your own models? | Opens when the Owner turns it on. After "Checking what this would stop…", it lists These agents stop answering (with each agent's chat model), These agents answer only from their knowledge base, No model of your own, so these stop:, Turned off for agents (with the agents using each service) and Turned off for the whole organization:, or "Everything already runs on your own models." Cancel or Turn on. Turning it off needs no dialog. |
| While on | Tasks without a model of your own read Off · no model of your own; Voice and the other services read Off · runs only on Qlar; Qlar models in the pickers are greyed out. |
Below it, the tasks come in groups. Every row shows the task, Sends: (what data the task gets), a chip for when it runs, and Runs on (the source of the model it uses now).
| Agent features | Conversation, Summarization, Personalization, Reranker, KB Training, Conversation Analysis. Leave one unset to inherit the platform default, shown as the placeholder ("smart · Platform default"). A choice whose model can't be used right now (subscriptions turned off, model removed) shows "alias is not available — using the Qlar default". Subscription models are not offered for KB Training. |
| Chat helpers | File check, Conversation wrap-up, Group chat attention, Read scanned PDFs, Escalation summary, Reaction follow-up, Bot detection. Read scanned PDFs offers only models that accept images and never uses the group's model; empty, Qlar's Azure AI Vision reads the pages. |
| Builder assistants | Agent setup writer, Test writer, Test judge, Logo brief, Logo image. Logo image offers only image models and never uses the group's model. |
| Plugins | Chat Form, SQL Database Reader, Dynamic Table. The group note says Dynamic Table searches embed on your Knowledge search model and Public Documents always runs on Qlar. |
| Voice | Speech to text, Text to speech, Voice calls · Turn-by-turn and Voice calls · Natural conversation, each with its own picker of models of that kind (placeholder, for example, "Qlar text to speech (gpt-4o-mini-tts)"), never the group's model. Without one, the row says how to add one ("Add a speech-to-text model (such as gpt-4o-transcribe or whisper-1)…"). Voice calls · Turn-by-turn (Realtime model) also offers "Qlar realtime voice mini (gpt-realtime-2.1-mini)"; Voice calls · Natural conversation (GPT-Live model, placeholder "Qlar GPT-Live (gpt-live-1)") offers only gpt-live-1 models. Both are the organization default for agents whose Voice page leaves Voice runs on at Organization default. Each counts its style's agents by where their calls run (Qlar platform, Your subscription, Your OpenAI account, Your Azure OpenAI account), with Show agents listing each agent, for example "Turn-by-turn · your OpenAI account (my-realtime) · organization default". See Voice, Plugins and Services on Your Own Models. |
| Other services | Knowledge search has an Embedding model picker ("Qlar embeddings (OpenAI)" when empty) that offers only embedding models, the note "Changing it rebuilds every agent's knowledge index on the new model: about n knowledge chunks, one call to your model per chunk.", and one row per agent: Indexing x of y, Ready · model, Rebuild failed (reason on hover) or Not indexed yet, with Retry next to a failed rebuild (or one without progress for half an hour) and Rebuild all under the list, which asks first. See Knowledge Search on Your Own Model. Web search has a Search account picker: Qlar's Exa account or Your Exa account · name for each service key. Public Documents has no picker: "Always runs on Qlar platform". |
Chat helpers, Builder assistants and Plugins each have a Use for every task in this group picker. A task without a model of its own uses the group's model (its placeholder reads "alias · from group"), then the Qlar default. Clear a task's picker to make it follow the group again; clear the group's picker to put its tasks without a model back on the Qlar default.
| Model list | Models switched on for the organization, grouped by source; a subscription model reads "alias · computer name". |
| Needs image input | Shown on File check. Only models that accept images are offered. "None of your own models can read images yet." when no API-key or subscription model can. |
| Personal agents only | Shown next to a subscription model. Subscription models never serve Service agents: in their chats the task uses the next model in the list. |