Skip to content
Ryvion

▪ The platform

Run, stream, verify.One platform.

An OpenAI-compatible inference gateway over a heterogeneous, verifiable fleet. Every output ships with a signed receipt. Every receipt resolves to lineage. Every lineage anchors to hardware that anyone can re-run.

Create workspace See pricing

▪ Capability stack

Eight surfaces. One verified runtime.

capability 01

OpenAI-compatible inference gateway

A drop-in /v1/chat/completions endpoint over the verified compute network. Same SDKs, same JSON, same streaming protocol — but every request is routed across heterogeneous Mac, Windows, and Linux operator nodes.

  • OpenAI SDK works unchanged
  • Streaming SSE + non-streaming
  • Per-key spend & rate limits
POST /v1/chat/completions
curl https://api.ryvion.ai/v1/chat/completions \
  -H "Authorization: Bearer ryv_…" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ryv-llama-3-8b",
    "messages": [{"role":"user","content":"Hi"}],
    "stream": true
  }'
HTTP 200 · text/event-streamopenai-compat

capability 02

Ghost-stream UX

Provisional tokens render ghosted while the network speculates. On commit they snap solid. On rollback they wipe and the replacement enters from the left. No other inference network exposes this.

  • Provisional → commit transition (180ms spring)
  • Rollback wipe with replacement
  • Built on multi-token speculative decoding
stream · ryv-llama-3-8bLIVE
The audit trail resolved cleanly across five nodes; receipts are signed and the lineage matches.Each provisional token will commit unless…

31.4

tok/s

2.1×

spec-up

142ms

ttft

capability 03

Receipts on every output

Every completion produces an Ed25519-signed receipt over output hash, model id, model revision, node id, timestamp, and (if audited) FuzzyTrace digest. Content-addressed CIDs you can verify offline.

  • RYV1 envelope · Ed25519
  • TranscriptDigest CID per output
  • Public key bundle for offline verify
RYV1 · receiptPASS
{
  "model": "ryv-llama-3-8b",
  "node":  "n_de_rtx7900_01",
  "tokens": 412,
  "output_hash":
    "sha256:9c1e4d7f42a8b3c1…",
  "transcript_digest_cid":
    "ryv1:sha256:e3b0c4429…",
  "signature":
    "ed25519:5c8f7a13b9d2e4f6…",
  "timestamp": "2026-05-12T18:42:09Z"
}

capability 04

Cross-node verification

Opt-in dual execution on Mac arm64 and Windows amd64 nodes. FuzzyTrace numerical comparison classifies every audited output as PASS, AMBIGUOUS, or FAIL — calibrated tolerance, not a hash match.

  • Dual-execute on heterogeneous hardware
  • FuzzyTrace numerical audit
  • PASS / AMBIGUOUS / FAIL verdict

FuzzyTrace audit · 2-of-2

M4 Pro
macarm64 · de · n_a1b2…
PASS
RTX 4070Ti
winamd64 · ca · n_c3d4…
PASS

0.0034

∆ logit

99.7%

overlap

PASS

verdict

capability 05

Heterogeneous fleet routing

Capability-typed router targets Mac M-series, NVIDIA RTX 30/40-series, and AMD RX 7000 nodes. Jurisdiction-aware. Reputation-gated. You see exactly which hardware ran your job in the receipt.

  • Mac arm64 · Windows amd64 · Linux amd64
  • Country-routed when policy demands it
  • Reputation tiers gate workload class

Online nodes · 6

+2 last 5m
M4 Pro
macus
RTX 4070 Ti
winca
RX 7900 XTX
linuxde
M3 Max
macuk
RTX 3090
linuxfr
RTX 4090
winca

capability 06

Knowledge & RAG

Upload documents, embed once with bge-m3, retrieve at inference time. Retrieved chunk hashes ride into the receipt so audits can reproduce the exact context window.

  • Verified embedding lane
  • Chunk hashes in the receipt
  • OpenAI Responses-API style retrieval

Knowledge · ks_org_eu_policy

policy.pdf142bge-m3
handbook.md38bge-m3
regs-2026.txt91bge-m3
retrieved 3 chunks · receipt embeds chunk hashes

capability 07

Agent hosting

Deploy long-running agents that share the same gateway, the same receipts, the same firewall. A managed runtime per agent — invocations are receipted just like raw inference.

  • Per-agent memory + tools
  • Receipts on every invocation
  • Optional PII firewall
agent_support_triageRUNNING
modelryv-llama-3-8b
memorykb_org_support
firewallstrict (PII redact)
invocations · 24h1,402
receipts · 24h1,402
runtime cost$0.21/hr

capability 08

Spend control & billing

Per-key budgets, alert thresholds, and a real-time spend ledger. Stripe-backed. No surprise bills — the ceiling stops requests, not your card.

  • Hard caps + alert tiers
  • Stripe Connect payouts to operators
  • Receipt-anchored audit trail

Spend · this month

$1,284.07

limit $5,000.00 · 25%

inference$842.10
embeddings$112.30
agent hosting$329.67
alerts50%, 75%, 100%

▪ What you don't get

Positioningmatters.

Inference networks blur into each other. Here's what Ryvion deliberately is not.

×

Not a GPU marketplace

We don't auction your job to whichever bidder shows up. The router picks based on capability, jurisdiction, and reputation — buyers don't negotiate with operators.

×

Not an OpenAI router

We don't proxy to OpenAI, Anthropic, or any closed provider. Models run on the verified Ryvion fleet. Same API surface, different substrate.

×

Not raw GPU rental

You don't get SSH. You don't manage CUDA. We give you a verified inference primitive — the platform handles fleet, drivers, and runtime.

▪ Workspace

Your dashboard, the moment you sign in.

Live receipts, live spend, live fleet. The buyer surface is built for the same audience as the receipts — people who want every cell to be a real value.

app.ryvion.ai/buyer/overview
▪ illustrative

▪ overview

Today

12 May 2026 · 18:42 UTC

requests

14,392

+8%

receipts

14,392

100%

p50 ttft

142ms

-12%

spend

$84.21

+4%

Recent receipts

streaming
PASSryv1:sha256:9c1e4d7f…a8b3ryv-llama-3-8b2s ago
PASSryv1:sha256:42e1c803…77f2ryv-llama-3-8b4s ago
PASSryv1:sha256:b81f0904…01c5bge-m36s ago
PASSryv1:sha256:1fa2b3c4…d5e6phi-49s ago
PASSryv1:sha256:7e80af11…3320ryv-llama-3-8b11s ago
NETWORK 6/6 ONLINETPS 32.4P50 142MS▪ ALL VERIFIED

▪ Start

Create a workspace, send your first verified request.

Create workspace See pricing