New features
Point an alias at an inference pool. An alias can now target an inference pool as well as a deployment. On Aliases, the target picker lists Deployments and Inference pools, the table shows each alias’s Target, and you can re-point an alias between a deployment and a pool without changing its URL. A pool alias needs an account-wide API key, serves the OpenAI-compatible routes, and addsGET .../v1/status and DELETE .../v1/session. Only the pool’s owner can point an alias at it. See Point an alias at an inference pool.Updates
- Alias tools in MCP. New
list_aliases,create_alias,update_alias, anddelete_alias. See MCP tools.
New features
Always on, from the endpoint page. Turn Always on on or off on any endpoint, or tick Keep always on when you deploy. The confirmation shows the hourly price and the cost per month. An always-on endpoint is never stopped for being idle, skips stop schedules, and comes back on a new GPU if it loses its own. A Monthly cost alert warns when its estimated monthly cost passes a threshold you set. See Always-on and idle stop.Know why a deployment failed. A failed endpoint shows what went wrong and a Next step, such as deploying again or picking a bigger GPU. A new Diagnosis card puts the latest machine reports and the stored logs in one place, and a starting endpoint shows its download progress and a boot timeline. See Deployment status and failures.See the cost before you deploy. The deploy form estimates the hourly cost (GPU plus storage), how long the endpoint will take to get ready, and what its disk costs per day while paused. See Deploy a model.Interruptible capacity. Deploy on cheaper GPU time that can be taken back when someone outbids it. An outbid endpoint shows Interrupted (outbid) and comes back on the same URL with Restart on a new instance. Fine-tunes started this way, with Interruptible (cheaper bid box) on New fine-tune or the AI assistant’s card, or from the API or MCP, resume from their last checkpoint. See Interruptible capacity.Inference pools. Pools on the Endpoints screen shows each pool, its members, their health, and the request slots in use. From the MCP server, manage pool members, size a managed pool, and run worker tasks on a pool. See Inference pools and MCP tools.Create a key from the endpoint page. Create a key for this endpoint shows the key once with a ready-to-paste config for Opencode and any OpenAI-compatible client: base URL, key and model name. It works while the endpoint runs, is paused, or is off on its schedule, and the API keys screen now offers those endpoints too. See Create a key for an endpoint.See when an endpoint runs. An endpoint stopped by its schedule reads Off (schedule) with the time it starts again, not Terminated; one stopped for being idle reads Stopped (idle), and one paused for balance or a spend limit Stopped (budget). Each endpoint shows its schedule in one line on the Endpoints list and at the top of its page, and its URL as soon as the model is running. See Deployment status and Schedules.Know a fine-tune is covered before it starts. New fine-tune and the AI assistant’s card show your Balance after the run. A run with an estimate starts only when the wallet holds the estimate’s high end plus a 20% buffer; otherwise Start fine-tune is disabled with the amount needed and an Add funds link. The API and MCP refuse it withbalance_below_estimate. See Before a fine-tune starts.Checks and targets. Under Test it on, pick checks suggested from your examples, such as Valid JSON, Matches the schema, Label accuracy, Number within 10% or an answer time, and a target for each. The version page shows a Checks table with each target Met or Missed, and Home shows N of M targets met on the model’s card. See Checks.PDFs and changes to saved data in the AI assistant. Attach a PDF with selectable text and the assistant builds examples from it (scanned PDFs have no text to read). Ask for changes to saved training data, such as “shorter answers” or “drop the long ones”, and the assistant shows them on a card and saves them as a new version of the dataset when you approve. See AI assistant.Small, Medium or Large hardware. Both deploy pages ask for a size instead of a GPU, with a price range per hour for each. Valkyrie rents the cheapest card in that size that holds the model, and Choose a specific card still opens the full list. See Hardware sizes.Invite people when you share a model. Share a model with an email address outside the workspace and Valkyrie sends an invitation. They get access to the model when they join. See Share a model.Updates
- Typical run on each model card. The models under Ready to fine-tune show the typical time and cost for 1,000 rows on the Standard recipe. See New fine-tune.
- Idle endpoints stop after 2 hours. An endpoint that serves no requests for 2 hours is stopped, unless it’s always on. The Always on run plan is now called Until stopped. See Idle stop.
- Paused endpoints bill their disk. The kept disk is billed while paused, and an endpoint paused for more than 72 hours is released. The endpoint’s page shows the daily cost and when the pause ends. See Paused endpoints.
- Scheduled stops release the GPU again. A scheduled stop now releases the machine, and a scheduled start rents a fresh one.
- Disks sized from the model. A fine-tuned version deploys with a disk sized from its trained weights. See Disk size.
- Model names checked before renting. A name that isn’t a Hugging Face
owner/nameid is refused before a GPU is rented. See Model names. - Every serving engine in the deploy form. The Serving engine list now shows every engine Valkyrie supports. See Serving engines.
- Wallet links. The Wallet lives at Settings > Wallet. Old links, including the return from a Stripe payment, still work. See Wallet.
- Invite roles capped. Nobody can invite someone as Owner. Owners and Admins invite as Admin or below; anyone else sharing a model can invite only as User or Viewer. A role you can’t grant is refused with a 403 and nothing is shared. See Invite people.
- Eval checks from MCP.
create_fine_tuning_jobtakeseval_checks, and the newsuggest_eval_checksproposes them for a dataset. See Eval checks. - MCP tools. New
set_deployment_always_on,set_deployment_idle_stop, anddiagnose_deployment, plusalways_on,idle_stop_minutes,capacity, anddry_runon deploys. See MCP tools. - Pay as you go with your own idle timeout. Pick how long an endpoint waits without requests before it stops, from 30 minutes to 24 hours, when you deploy or later on its page. The page shows when it will stop and offers Start again after an idle stop. See Set your own idle timeout.
- Compare a fine-tune with its adapter off. In the Playground, a LoRA fine-tune compares against This endpoint, without fine-tuning by default, with no second GPU, and each answer shows its response time. See Compare side by side.
Bug fixes
- Spend limits pause endpoints reliably. Some endpoints kept running after their workspace or team passed its cap, or ran out of balance.
- Stalled downloads fail fast. A model download with no progress for 5 minutes fails the deployment and releases the machine, instead of billing while it waits.
- No more endless loading. A dashboard request that gets no answer now says so and lets you try again.
- The dashboard works at phone width. Navigation, sign-up, and the deploy form no longer scroll sideways.
New features
Workspaces and teams. Create a workspace for your organization from the switcher in the top bar, invite people by email with a role, and group them into teams. Pending invitations appear in the switcher. Manage members, teams, and invites under Settings.Share models with people and teams. A model’s Access tab grants members or teams use or manage. Removing someone revokes their keys for that model.Choose who owns new work. New fine-tunes and endpoints can belong to you, a team, or the workspace.Monthly spend limits. Set monthly caps for the workspace and each team under Settings > Billing. Owners and billing members get an email at 80% and 100%. A deployment that crosses a cap pauses, like a low balance. See Wallet & billing.Always-on endpoints recover on their own. If an always-on endpoint loses its GPU, Valkyrie starts it on a new one, retrying up to 3 times.Tune every training setting. New fine-tune now lets you edit the optimizer, max sequence length, and seed, and offers all seven learning-rate schedulers. The time and cost estimate follows batch size and sequence length. In the AI assistant, ask to change any setting, or switch the base model from the card. See New fine-tune and AI assistant.Updates
- One wallet per workspace. Your balance and top-ups follow the active workspace, and billing moved to Settings > Billing.
- Scheduled stops pause instead of delete. The GPU stops billing but keeps its disk, so scheduled starts are faster. See Operate an endpoint.
- Clear status for starting endpoints. MCP deployment status now says what a starting endpoint is doing, such as downloading weights with progress or loading the model. See MCP tools.
- Stop schedules on always-on endpoints are rejected. They could never run.
Bug fixes
- vLLM endpoints start up to 15 minutes sooner. New deployments and resumes no longer wait idle before serving.
- Resumed endpoints serve reliably. An endpoint that doesn’t serve after a scheduled start is moved to a fresh GPU instead of failing.
- Paused endpoints show as paused, not failed or unhealthy.
- No charge for paused time. Billing for a resumed endpoint starts when it resumes.
- Top-ups always credit your wallet. Some early accounts had no wallet, so a payment could fail to credit.
Bug fixes
- Faster endpoint starts. Your model starts serving sooner once its GPU is ready. See Deployment stages.
- Stuck endpoints fail fast. If a GPU never starts your model, Valkyrie stops it after about 10 minutes and asks you to deploy again, so you aren’t billed for an idle machine.
- Base models land on a GPU that fits. Large base models no longer start on a GPU with too little memory and stall. See Deploy a base model.
- Accurate endpoint health. An endpoint that is still loading its model is no longer reported as healthy.
New features
Check your data before it’s saved. The AI assistant now shows you what it read from your table and the training examples it built, before it saves anything. Click Looks right or Something’s off. Title rows, repeated headers, and TOTAL rows in messy spreadsheets are skipped automatically. See AI assistant.See each step of a fine-tune starting up. A starting run lists its steps (renting a GPU, GPU up with its rate, script started, loading the model, training) with a time for each. If the GPU is up but training hasn’t begun, a warning says when the run will stop so you aren’t charged for an idle machine. See While it trains.Select several at once. Delete multiple datasets on Data, or archive and unarchive multiple models on Models. See Data and Archive a model or version.Search your datasets on Home. The Use a dataset picker is now a search box by name or ID, showing each dataset’s row count. See Home.Updates
Number predictions scored by closeness. Fine-tunes that predict a number (a price, a count) report Within 10%, Within 20%, and Average error instead of exact matches. See Test results.Lower training costs. Each fine-tune now runs on the cheapest GPU with enough memory for its model.Tables in the assistant chat. Ask for a table, for example to compare your fine-tune with its base model, and it renders as one. Links in the chat open in a new tab, so you keep your conversation.Faster Data page. The dataset list loads quickly, even with many datasets.Bug fixes
- Fine-tunes no longer stall on “Starting”. Some runs never began training and were stopped after several minutes.
- Accurate run cost and time. A run’s page shows what it was billed, instead of $0.00, and finished runs show their training time.
- The assistant reuses training data as is. Picking a dataset the assistant already built no longer creates a converted copy.
- Dark theme is readable everywhere. Sign-in fields, dialogs, forms, and the error page no longer show light-on-light text.
- The chat survives a product update. If Valkyrie updates while a chat is open, you’re asked to reload, and your unsent message is kept.
- Requests to a busy deployment no longer hang. Streaming calls could block later requests to the same deployment.
- More reliable deployment starts. Deployments with long model downloads no longer fail at start-up, resuming reports clearly when the GPU is unavailable, and deployments pinned to a specific GPU offer no longer fail at random.
- Tool calling works on TabbyAPI deployments. These deployments now start serving on their own and return proper tool calls to agent clients. See Connect a model.
New features
New model or new version, your choice. When you start a fine-tune, choose New model or New version of a model. A version can use different examples or another base model, and version numbers no longer depend on the dataset’s file name, so retraining stays under one model as v2, v3. The AI assistant’s card says which model and version it will be, and Change model switches it. Attaching the same file again reuses its dataset instead of adding a copy. See Models and New fine-tune.Name your models, and rename them. A fine-tune can have a name instead of its dataset’s file name: set Model name in New fine-tune, or let the AI assistant name it after what it does. Rename on a model’s page renames every version. See Rename a model.Updates
Results arrive in the chat. When a fine-tune you started with the AI assistant ends, its result is posted to that conversation, even if you left, with a suggested next step when the gain is small. See AI assistant.See and change the settings before you confirm. The assistant’s card now shows the name and the passes over the data, and you can change either by telling it. See You stay in control.Time left shows a time. A run’s page counts down from its estimate from the start, instead of showing “Estimating…” until training is nearly done. See While it trains.Bug fixes
- A just-started fine-tune is no longer reported as ready. The assistant could say a fine-tune was ready, and show its check as failed, right after you started it.
New features
See how much better your model got. New fine-tunes hold back part of your examples, then answer them with the base model and with your version. The version page shows the gain in points, answer match, exact answers and loss for both, and a few answers side by side. Choose how much to hold back under Test it on. See Test results.Time and cost before you start. New fine-tune now estimates how long a run takes and what it costs, from your row count, the model’s measured speed, and current GPU prices. See New fine-tune.Training loss chart. Every version draws its training loss over the run, live while training and on finished versions. See Training loss.Choose when an endpoint runs. Deploy pages ask When should it run?: Always on, or Office hours on weekdays between the hours you choose, each with its monthly cost. See Deploy a fine-tuned version.Compare two endpoints side by side. The Playground sends each message to two endpoints at once, for example your fine-tune and its base model. See Compare side by side.Example datasets. Copy a ready-made set of about 2,000 real examples (support replies, ticket triage, invoice to JSON, plain English, general instructions) to try a fine-tune before you have your own data. See Start from an example set.Clean up a dataset. Remove duplicates or over-long examples in one click. Each cleanup saves a new version and keeps the original. See Clean it up.AI assistant (preview). Describe the task and give it your material, and it builds the training set and fine-tunes for you. It is rolling out gradually and shows a padlock where it isn’t available yet. See AI assistant.Updates
The AI assistant does the work. Attach a spreadsheet (Excel, CSV, JSON, Parquet and more) and say what to predict, or give it text to learn from. It prepares the training data, picks a base model, starts the fine-tune after you confirm the cost, reports how much better it got, and deploys it. See AI assistant.One box on Home. Write what you need once, then choose the AI assistant or your own examples in the same box. See Home.Every row counted, and a preview. Dataset pages now show the total number of rows and the first examples as training reads them. See A dataset’s page.Clearer model choices. Recommended is now Ready to fine-tune, with size, memory, what each model is good for, and a Tested badge on models fine-tuned on Valkyrie. Qwen2.5 3B Instruct joins the list. See New fine-tune.Bug fixes
- Latest loss shows again. Training runs on recent library versions reported no loss in the dashboard. It now appears, along with the full chart.
New features
A dashboard built around fine-tuning. The dashboard now follows the path from your data to a live model. Home asks what your model should get better at and starts a fine-tune in one go. Models groups every training run into versions (v1, v2, and so on), with a Pipeline view that shows where each model is. Data checks the format of every file and shows which model versions it taught. Endpoints lists the models you can call, and Playground lets you chat with any of them. See the dashboard overview, Models, Data, Endpoints, and Playground.A guided deploy page for fine-tuned versions. Click Deploy on a completed version to pick the hardware from a list with live GPU prices and memory per card, choose the engine, and see the cost per hour before you deploy. See Deploy a fine-tuned version.Archive and unarchive models. Archive a single version or a whole model to tidy your list. Nothing is deleted: outputs, billing, and usage history stay, and Show archived brings them back. See Archive a model or version.Updates
Base model or Fine-tuned, at a glance. Every endpoint now carries a Base model or Fine-tuned pill, wherever it appears. See Endpoints.Deploy a base model from anywhere. Deploy a base model is now reachable from Home, Models, and Endpoints. See Deploy a base model.Deployment key scopes list only active deployments. When you create a deployment key, the scope list offers only deployments that are pending, starting, or running, plus Account-wide. See API keys.Quota is now Wallet. The Quota item in the sidebar is now called Wallet. Your balance, charges, and top-ups work the same way. See Wallet.Clearer start-up stages. The “Downloading weights” stage now reads Preparing: the machine is up and the runtime and model files are being set up. If a deployment is still starting after 25 minutes, the dashboard shows Taking longer than usual. See Deployment stages.Bug fixes
- No GPU left billing after a failed start. When a deployment cannot start, Valkyrie now cleans up the failed attempt before retrying, so it no longer leaves a GPU running and billing.
- Unreachable machines stop on their own. A deployment whose machine never becomes reachable now stops automatically, with a clear error explaining why. See Deploy a model.
- Cancelling a starting deployment sticks. A deployment you cancel while it is still starting no longer comes back as Preparing.
- Fine-tuned endpoints answer whatever model name you send. Calls to a fine-tuned endpoint’s API URL no longer return a 404 when the request names the base model. See Call your deployment.
Bug fixes
- No extra charges after a restart. Restarting a deployment (manually, on a schedule, or by automatic recovery) no longer adds an “Instance usage” charge for the time it was stopped. Valkyrie now bills only from when the deployment’s current GPU starts provisioning. See Quota & billing.
Updates
Card top-ups are now live for everyone. Self-serve top-ups through Stripe are now available in production. Add funds from the Quota screen at any time. Your balance updates once Stripe confirms the payment. See Adding funds.Bug fixes
- Cleaner top-up confirmation. The confirmation screen after a successful top-up no longer shows an outdated reference code or caveat.
New features
Top up your balance by card. Add funds to your Quota balance yourself, any time, right from the dashboard. Pick or enter an amount, pay on Stripe’s secure checkout, and your balance updates automatically once the payment is confirmed — no need to request a top-up and wait. See Adding funds.Page through MCP logs.get_persisted_logs now returns stored logs one small page at a time, newest first. Pass the returned next_before_id as before_id to fetch older entries. MCP log tools also cap output at 500 lines, so large logs return quickly instead of timing out. See Paging through persisted logs.Updates
Send your MCP key in a request header. Remote MCP clients now authenticate with anX-API-Key or Authorization: Bearer header on every request. If you also pass a key in the URL, the header wins. Remote requests without a key no longer run under a shared server key, so each request acts only on your account. See MCP setup.Bug fixes
- Base-model deploys from the dashboard work again. Deploying a preset from Models no longer fails right away with “Extra inputs are not permitted”. Models without a supported launch preset now show a clear message before you submit. See Deploy a model.
- No duplicate GPUs from scheduled starts. A scheduled start or automatic restart no longer leaves extra GPU instances running when a machine fails to start. See Managing a deployment.
- Failed deployments release their GPU. Valkyrie now finds and shuts down GPU instances left running after a deployment fails.
- Balance checks cover every running GPU. A deployment that leaves the Running state while still holding a GPU is now included in Valkyrie’s spending checks. See Quota & billing.