Skip to content

Ryvion V8 · live network

The verification layer for distributed AI compute.

Ryvion routes inference across named hardware and returns a signed, browser-verifiable receipt for every output. Ghost streaming and speculative decoding make the path faster without hiding uncertainty.

Avg speedup
2.82×
reference bench
Receipt CID
100%
gateway target
Hardware
3
vendors verified

▪ claims stay tied to receipts and live network metrics below

Live demo

phi-4 · eagle-2RTX 4070 Ti Super
user >What is the capital of France?
assistant > 
tps · decode
0.0
spec accept
0%
tokens
0 + 0p
x-ryvion-receipt-cidryv1:sha256:9c1e4d…a7f3c9d8

▪ animation runs client-side · same protocol as production stream

Live network

ONLINE0/0REGIONS0RECEIPTSwaiting for persisted V8 runsONLINE0/0REGIONS0RECEIPTSwaiting for persisted V8 runs

Network telemetry

Real numbers from the production network

○ connecting

Online nodes
0operators
live from marketplace
Jurisdictions
0regions
active fleet coverage
Tokens streamed
0rolling
across last 32 runs
Speculative runs
0of 32
drafter accepted
TPS p50
0.0tok/s
decode median, last run
Default score
0.95baseline
honest operator floor

Trust layer

The receipt is the product.

Every output emits a content-addressed dependency tree. Six receipts, one chain: prompt → policy → plan → assignment → transcript → signed result. Each node is an Ed25519-signed CID anyone can verify.

  • Deterministic CIDs — same inputs, same hash, every time
  • Ed25519 signatures by the executing node
  • Verifiable from any browser — no API key, no auth
  • EU AI Act Article 14 evidence-ready
▪ V8 receipt lineage · illustrative
PASS
01
01 · INGRESSPromptRoot

hash binding · gemma-4-26b-a4b-it

parent of all receipts in this trace

02
02 · POLICYPolicy

max_tokens, temperature, sla · backend=llama-server

03
03 · PLANExecutionPlan

role=target_verifier_consumer · capability-typed

04
04 · ASSIGNAssignment

node 0e5125f14e7e…a71bc7 · RTX 4070 Ti Super

05
05 · EXECTranscriptDigest

739 tokens · finish=stop

06
06 · RECEIPTTargetReceipt

signed by 0e5125f14e…a71bc7 · Ed25519

07
07 · AUDITFuzzyTraceoptional

not sampled · stochastic audit (p≈0.05)

dependency tree · content-addressed · Ed25519 signedVerify any CID at /v8

▪ Heterogeneous fleet

Real hardware. Named.

Inference fleets are usually a black box of “GPU instances”. Ryvion shows the SKU, jurisdiction, vendor, and reputation tier of every node. Mac arm64, NVIDIA RTX, AMD RDNA — we verify across architectures so the answer doesn't depend on whose hardware happened to win the bid.

▪ Fleet snapshot

4 of 5 nodes online · 3 jurisdictions · Mac arm64 + NVIDIA + AMD

  • DE

    ryv-de-pri-01

    RX 7900 XTX · 24GB

    trusted● online
  • CA

    ryv-ca-north-04

    RTX 4070 Ti Super · 16GB

    elite● online
  • FR

    ryv-fr-lyon-02

    RTX 4060 · 8GB

    trusted● online
  • CA

    ryv-ca-cpu-07

    CPU node · 16GB

    trusted● online
  • DE

    ryv-de-prv-09

    RTX 3090 · 24GB

    watch○ offline

▪ Developer path

Switch the base URL. Keep the receipt.

The OpenAI SDK works untouched. You get streaming, tool-use, jurisdiction routing, and a signed receipt header — all without learning a new client.

▪ python · openai sdk 1.xbase_url=api.ryvion.ai/v1
from openai import OpenAI

client = OpenAI(
    base_url="https://api.ryvion.ai/v1",
    api_key="ryv_live_...",
)

stream = client.chat.completions.create(
    model="phi-4",
    stream=True,
    messages=[
        {"role": "user",
         "content": "Why distributed inference?"},
    ],
    extra_body={"jurisdiction": "CA"},
)

for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

# Receipt CID is on every response:
#   X-Ryvion-Receipt-CID: ryv1:sha256:9c1e4d7f...
# Verify any CID at https://ryvion.ai/v8 — no auth required.
▪ stream · phi-4ssl · http/2
$ curl https://api.ryvion.ai/v1/chat/completions \
-H 'Authorization: Bearer ryv_live_…' \
-H 'Content-Type: application/json' \
-d '{"model":"phi-4","stream":true,
"messages":[{"role":"user",
"content":"Why distributed inference?"}]}'
data: Distributed inference
data: lets you
data: verify every
data: output against
data: the hardware that
data: produced it.
data: [DONE]
▪ response headeraudit=PASS
x-ryvion-receipt-cidryv1:sha256:9c1e4d7f…a7f3c9d8
Verify this CID

▪ Versus the network

Same API. Different premise.

Most networks ask: how cheap can we rent the GPU? Ryvion asks: can you prove the answer? The OpenAI-compatible surface is the same. The artifacts you take home are different.

▪ vs network
Ryvion V8
the verification layer
Hyperbolic
marketplace
OpenRouter
model router
io.net
GPU rental
Signed receipts on every output
Ed25519 + RCOG
Speculative streaming with rollback
provisional → commit
Cross-node verification
FuzzyTrace audit
Open hardware visibility
typed fleet snapshot
GPU model only
operator-side
OpenAI-compatible API
limited
Heterogeneous (Mac · NVIDIA · AMD)
NVIDIA only
Measured speculative speedup
receipt-backed metrics

Ship today

Send the first request. Verify the receipt.

Create a workspace, point your SDK at api.ryvion.ai/v1, and inspect the lineage of every output. The free tier ships with speculative acceleration and signed receipts on day one.