We operate the hardware ourselves. You get direct pricing, predictable performance, and full visibility into the model you run.
100,000 free tokens. No card required. A 6-digit code is emailed; your key is shown here once.
It's already pasted into the request on the right. Run it.
curl https://api.earthruntime.com/v1/chat/completions \ -H "Authorization: Bearer $EARTHRUNTIME_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "qwen3.6-35b", "messages": [{"role": "user", "content": "Explain why tail latency matters for production inference in three sentences."}] }'
curl https://api.earthruntime.com/v1/chat/completions -H "Authorization: Bearer $EARTHRUNTIME_KEY" -H "Content-Type: application/json" -d '{"model":"qwen3.6-35b","messages":[{"role":"user","content":"Explain why tail latency matters for production inference in three sentences."}]}'
Works with your existing OpenAI client — change the base URL and the key, keep everything else. Read the docs →
Every model is served from hardware we own and operate. What you see on the card is what runs — precision, context window, and whether reasoning is on.
Served from Qwen's own official FP8 checkpoint. We do not requantize it ourselves. Reasoning is disabled by default for lower latency; enable it when the workload benefits.
The smaller Qwen for cost-sensitive, high-volume work.
reasoning_effortOpenAI's open-weight model, served without a middleman.
Not yet available; the rate above applies at launch. Email contact@earthruntime.com to be told when it opens.
Measured from a neutral vantage outside our network, at 64 concurrent requests, on 2026-09-24. The exact script, workload and error definition are on the methodology page. Run it against us and against anyone else.
| Model | Throughput | TTFT p50 at c=64 | Latency p50 | Latency p95 | Errors |
|---|---|---|---|---|---|
qwen3.6-35b | 2,464 tok/s | 765 ms | 2.47 s | 3.56 s | 0% |
gpt-oss-120b | 2,276 tok/s | 537 ms | 3.11 s | 4.57 s | 0% |
qwen3.8-27b | 1,608 tok/s | 508 ms | 4.70 s | 8.59 s | 0% |
| Zero errors across all 3,104 requests. TTFT is quoted at 64 concurrent streams, which is where it was measured; at a single stream it is 73 ms. Why the gap. | |||||
| Endpoint | Throughput | Latency p50 | Latency p95 | Errors |
|---|---|---|---|---|
| earthruntime | 2,464 tok/s | 2.47 s | 3.56 s | 0% |
| OpenRouter, default routing | 1,534 tok/s | 1.65 s | 4.68 s | 0% |
| OpenRouter to AtlasCloud | 1,327 tok/s | 1.87 s | 3.18 s | 0% |
| OpenRouter to Parasail | 3,042 tok/s | 2.18 s | 2.74 s | 17.6% |
| Read those two right-hand columns together. Parasail posts the highest raw throughput because it returns 429 to 17.6% of the load and finishes a smaller job sooner. OpenRouter's default routing is genuinely faster than us on the median; we are ahead on throughput, on the p95 tail, and on serving every request we were sent. | ||||
If your code already talks to OpenAI, it already talks to us.
Enter your email at the top of this page. A 6-digit code arrives, you paste it, and the key is shown once — with 100,000 free tokens already on it.
Set base_url to https://api.earthruntime.com/v1 and pass the key. Request and response shapes are unchanged, streaming included.
Use qwen3.6-35b, qwen3.8-27b or gpt-oss-120b. Nothing else in the call changes.
from openai import OpenAI client = OpenAI( base_url="https://api.earthruntime.com/v1", # step 2 api_key=os.environ["EARTHRUNTIME_KEY"], ) stream = client.chat.completions.create( model="qwen3.6-35b", # step 3 messages=[{"role": "user", "content": "Hello"}], stream=True, ) for chunk in stream: print(chunk.choices[0].delta.content or "", end="")
import OpenAI from "openai"; const client = new OpenAI({ baseURL: "https://api.earthruntime.com/v1", apiKey: process.env.EARTHRUNTIME_KEY, }); const r = await client.chat.completions.create({ model: "gpt-oss-120b", messages: [{ role: "user", content: "Hello" }], }); console.log(r.choices[0].message.content);
Every rate on this page is plain HTML — nothing is loaded by script. The same numbers are published as /pricing.md for machines.
| Model | Input / 1M | Output / 1M | Context | Notes |
|---|---|---|---|---|
qwen3.6-35b | $0.10 | $0.90 | 262K | Qwen official FP8; reasoning off by default |
qwen3.8-27b | $0.24 | $2.20 | 262K | Qwen official FP8; reasoning off by default |
gpt-oss-120b | $0.03 | $0.17 | 131K | MXFP4; reasoning on by default |
deepseek-v4-flash | $0.30 | $0.90 | 262K | Not yet available; rate applies at launch |
Rates are per million tokens and are billed on the usage block of each response. Notice periods for rate changes are in the terms. | ||||
100,000 free tokens for new users. No card required. Get a key at the top of the page — it arrives with the tokens already on it.
Need more than the packs cover, or a committed rate? contact@earthruntime.com.
Credits are added to the key tied to the email you enter. No subscription, no account setup.
Checkout is handled by Stripe. Token estimates assume a typical input/output mix.
A marketplace resells someone else's GPUs and takes a margin on every token. We operate inference hardware in cities and look for local uses for its physical outputs — so the price you pay is ours, and the performance you see is ours to keep.
Inference hardware sited in cities, run by us, no intermediary.
The first system pairs that compute with direct-air carbon capture.
Recovered CO2 is bottled and supplied to nearby hospitality businesses.
Faster inference is the product today. A more useful relationship between cities and compute is what we're building toward.
Yes: the carbonation in your soda at a local bar can come from the machine that answered your API call.
“There were no latency issues using Qwen.”
A developer tested Qwen on a Rust DSP mastering tool containing a deliberately misleading hardcoded ID. The model attempted the obvious fix, saw its tests fail, backtracked, identified the underlying cause, and produced a patch that two other models independently reviewed.
“Much better than the OpenRouter options we tested.”
After benchmarking earthruntime against OpenRouter-hosted alternatives in production, the team made earthruntime its primary provider and retained the marketplace as an outage fallback.
Qwen 3.6 35B (qwen3.6-35b), Qwen 3.8 27B (qwen3.8-27b), and GPT-OSS 120B (gpt-oss-120b). DeepSeek V4 Flash is coming soon.
Yes. Point your existing OpenAI client at https://api.earthruntime.com/v1 with your earthruntime API key; request and response formats stay the same.
Yes — 100,000 free tokens to start, no card required.
We operate the inference hardware ourselves, so you get direct pricing, predictable performance, and full visibility into the model you are running — no marketplace overhead.
Qwen 3.6 35B is served from Qwen's own official FP8 checkpoint, at its full 262K context. Nothing is requantized by us, and the precision each model is served at is listed on its model card.
No subscription. Buy a credit pack ($1, $5 or $20) with the email address your key is tied to; credits land within a minute. For higher volumes, email contact@earthruntime.com.
In cities, on hardware we operate ourselves. Our first system pairs the compute with direct-air carbon capture and supplies the recovered CO2 to nearby hospitality businesses.