Enterprise performance, measured like it's already live
How requests move through the platform
Four layers between your HTTP call and a streamed token. Each one is boring, observable, and fast.
Latency-based routing across US regions. TLS terminated at the edge, rate limits applied before anything touches compute.
Priority queues per model. Load balanced across the healthy node pool, KV-cache aware placement.
Continuous batching, unbuffered token streaming, automatic prompt deduplication.
Dedicated Instinct MI355X nodes per workload. Node failure detected in <1s, traffic rerouted without dropping connections.
Dedicated silicon. No abstraction leaks.
Workloads land on dedicated AMD Instinct MI355X accelerators, tuned and monitored end to end — never oversubscribed, never sharing a box with someone else's training run.
What the platform does
Tokens stream as they're generated — no waiting for the full response. Fast first token, steady throughput.
We handle the scaling. Traffic spikes are absorbed automatically. You never think about capacity.
Latency-based routing across US regions. Requests hit the nearest healthy node.
KV-cache reuse across requests. Automatic prompt deduplication.
Node failure detected in <1s. Traffic rerouted without dropping connections.
An inference company, not an AI lab.
We don't train models and we don't build products on your data. API requests are processed in memory and never written to disk — zero retention by architecture, not by policy. Conversations in our desktop and web apps are kept for your own history only: never used for training, and yours to export or delete at any time.
API prompts, completions, and uploaded files live in volatile memory for the duration of the request. Then they're gone — ZDR by default, no exceptions.
No fine-tuning, no evals, no "service improvement" pipelines. Nothing sent through the API or the apps is ever used to train a model — ours or anyone else's.
Conversations in our desktop and web apps are stored so your history follows you. Never used for training, and yours to export or delete at any time.
We don't build or sell models, so your data is never feedstock for a competing product or a revenue stream. Dedicated enterprise inference under contract — no data resale, no training incentive.
Run the data layer in your VPC — or the whole platform on-prem.
For regulated teams and law firms, where data sits is the security boundary. Walrus — the SQLite-over-S3 store our apps and agents write conversations, uploads and generated artifacts into — can live entirely in your own environment, under keys namespaced by your organization id. Your infrastructure, your keys, your retention policy.
- On-prem or in your VPC. Run the full API, Postgres, Valkey and object store inside your perimeter; only inference requests and usage events leave.
- Walrus in your bucket. One SQLite database per user and per workspace, written under a lease in your Valkey to MinIO, Ceph, R2 or S3 — not ours.
- Only prompts cross the boundary. Inference goes through the basemnt Router, which never persists prompts or completions.
- Retention you control. Per-org retention windows; hard deletes remove the database row and its object in one sweep.
Drop-in apps your team already knows.
Polished desktop and web assistants that feel instantly familiar — wired to your models, your SSO, your audit logs. Migrate seats in an afternoon, not a quarter.
- Instantly familiar. Drop-in replacements for the AI chat apps your team already uses every day.
- SSO & SCIM provisioning. Seats follow your identity provider, not a spreadsheet.
- Your models, your rules. Per-workspace model selection, admin controls, usage caps.
- Audit everything. Full request logs and cost attribution per seat, per team.
- Retained, never trained on. App conversations are kept so your history follows you — never used to train models, and yours to export or delete at any time.
{
"model": "GLM-5.2",
"messages": [{"role": "user", "content": "Hello"}],
"stream": true,
"temperature": 0.7
}
{
"id": "chatcmpl-...",
"model": "Kimi-K2.7",
"choices": [{
"delta": {"content": "Hello"}
}],
"usage": {"prompt_tokens": 12, "completion_tokens": 284}
}
Flat plans. Metered by requests, not tokens.
We don't count tokens — every plan bills on total requests, so prompt size never changes your price. ZDR is on by default for every API request.
- 500 requests / 5 hours
- 10k requests / week
- Multiple models, one endpoint
- Zero data retention (ZDR) on every request
- 1,500 requests / 5 hours
- 30k requests / week
- Multiple models, one endpoint
- Zero data retention (ZDR) on every request
- 3,500 requests / 5 hours
- 75k requests / week
- Multiple models, one endpoint
- Zero data retention (ZDR) on every request
- Custom request ceilings
- Dedicated capacity on MI355X clusters
- VPC / private deployment options
- Custom SLAs & compliance reviews
- SSO & SCIM provisioning
- Zero data retention (ZDR) on every request
Start building.
Deployment, volume, or a custom SLA — tell us what you're building and we'll follow up within one business day.