TACHYON·A / Overviewpublic · v0.4.0·doc / TX-0001
AoverviewBplaygroundCdocsDpricingEstatusFdashboard
Inference proxy / spec sheet

The fastest hosted LLM, exposed as a plain OpenAI endpoint.

Tachyon is an inference proxy that exposes the world's fastest hosted LLM — Taalas HC1 silicon running Llama 3.1 8B — as a plain OpenAI-compatible endpoint. No SDK changes. No custom integrations. Just point your existing OpenAI client at Tachyon and get ~17,000 tokens per second with sub-4ms time-to-first-token.

Generate API key →Read the docs
Datasheet
ModelLlama 3.1 8B
BackendTaalas HC1 (custom silicon)
Endpoint/v1/chat/completions
FormatOpenAI-compatible JSON / SSE
StreamYes (SSE)
AuthNone required
Regionus-east
Price$0.05 / Mtok (in & out)
p50 ttft · us-east
3.4ms
peak throughput / user
17,000tok/s
uniform — in & out
$0.05/ Mtok
vs next fastest provider
≈10×
§ 01Signal path
YOUR APPTACHYONHC1 SILICON~0.3ms~3.0msRETURN · SSE TOKEN STREAM
Fig. 01 — Signal path

Your application sends a standard OpenAI-compatible request to the Tachyon proxy. The proxy forwards the payload to Taalas HC1 silicon, where Llama 3.1 8B is hardwired directly into custom hardware. Because the model weights live in silicon rather than DRAM, inference happens with negligible memory latency.

The full response is buffered on the HC1 and streamed back to your application as an SSE token stream. At ~17,000 tokens per second, the entire round-trip — including network overhead — completes in under five milliseconds. This is roughly an order of magnitude faster than GPU-based inference providers.

Note: Burst mode is intentional. HC1 throughput is so high that buffering the full response and sending it as one chunk is faster than incremental streaming for real-world response lengths.
§ 02Quickstart
Python — OpenAI SDK
from openai import OpenAI client = OpenAI(    base_url="/v1",    api_key="anything") response = client.chat.completions.create(    model="llama3.1-8B",    messages=[{"role": "user", "content": "Hello!"}])print(response.choices[0].message.content)
Shell — curl
curl -N -X POST /v1/chat/completions \  -H "Content-Type: application/json" \  -d '{    "model": "llama3.1-8B",    "messages": [{"role": "user", "content": "Hello!"}],    "stream": true  }'
§ 03Pricing — at a glance
Provider$/Mtok in$/Mtok outp50 ttftPeak tok/s
Tachyon · HC1$0.05$0.053.4 ms17,000
Groq · LPU$0.10$0.1050 ms800
Cerebras$0.15$0.15200 ms1,800
Together · H100$0.30$0.30120 ms90
OpenAI · gpt-4o-mini$0.15$0.60300 ms120
§ 04Working name — alternates
Tachyonworking
Mach
Photon
Tick
Hertz
Sub5

Stop waiting on tokens.