What your agents are actually talking to

Qwen3.8-27B is an open-weights model released by Qwen in August 2026 under Apache 2.0, with a 256K native context window (model card).

We serve 4/5-bit quantized weights so the model fits in GPU memory with room left for long contexts. Every server downloads the same pinned build and checks its size and SHA-256 before it serves a single request:

Qwen's published benchmarks are for the full-precision model, and a quantized build can score a little lower. That's why we're measuring our own endpoint: see quality below.

Our servers run NInfer, an open-source inference engine by Neroued. Credit where it's due: its work is a big part of why a 27B model can serve long agent sessions at a flat price.

Each server runs one copy of the model with up to 8 requests generating at once. The requests on a server share its KV cache memory, and each one can use up to 128K tokens of context, depending on your plan.

In front of the engines, our gateway checks keys and plan limits, runs a fair queue across everyone when servers are busy, and sends each conversation back to the server that already has it cached, so follow-up turns don't reprocess the whole prompt.

We rent dedicated GPU servers by the hour from GPU cloud providers and change how many run as demand changes. Servers are health-checked continuously and taken out of rotation if they stop responding. The providers we use are listed in the Privacy Policy.

Live capacity, meaning servers online, slots in use, requests waiting and queue wait over the last hour, is on the status page.

• First token: under a second for a short prompt, up to about 2.5 s when a server is full.
• Long prompts: a server reads prompts at ~700–950 tokens a second before it starts answering. A 19K-token agent context takes about 20 seconds on its first turn. After that, each turn of the conversation is served from the cache, so only the new part of the prompt is read.
• Generation: ~42 tok/s for a single stream, and ~29–31 tok/s each when a server is running 8 at once.
• For scale: about 61 tok/s for Claude Opus on Anthropic (OpenRouter's reported throughput, September 2026). We're slower; what you get instead is no meter.
• Busy times: a request that finds every slot busy waits in a fair queue instead of failing. A response reports its wait in the x-queue-ms header (a streamed response that waits more than about 10 seconds is kept alive with keep-alive events instead, and its headers go out before the wait is known).

Measured September 2026 with the production engine settings on the server itself (network time not included) and 400-token answers.

Qwen's own published scores, and where the model is and isn't a good fit, are on the homepage.

We never use your content to train models, and we keep request metadata such as token counts, timings and status for your usage dashboard.

The Privacy Policy has the full detail.

• GET /v1/models lists the model and the context window the engines report.
• Every response's model field names the model that answered, and its x-queue-ms header shows how long it waited.
• The status page shows servers online, slots in use and queue wait, live.
• Our promise about limits is in the pricing page and the Terms of Service.

# The model and the context window the engines report
curl -s https://api.nevermetered.com/v1/models -H "Authorization: Bearer $NEVERMETERED_API_KEY"

# Response headers: x-queue-ms is how long the request waited for a free slot
curl -s -D - -o /dev/null https://api.nevermetered.com/v1/chat/completions \
  -H "Authorization: Bearer $NEVERMETERED_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"qwen3.8-27b","max_tokens":16,"messages":[{"role":"user","content":"Hi"}]}'