Qwen 3.8 27B · FP8 weights and KV cache · OpenAI & Anthropic compatible
A guaranteed GPU share for your always-on agent.
Flat-rate inference with no token usage caps. Reserved capacity is guaranteed under contention. Additional usage available on a best-effort basis.
Live capacity
Qwen 3.8 27B
Live numbers are unavailable this second. The commitment behind them does not move: committed shares stay at or under capacity as exact integer arithmetic, checked under a database lock on every checkout — which is why tiers here sell out instead of getting slower.
Keep the SDK you already use
Two protocols, one endpoint. Change a base URL and keep the client, the types and the retries you have already debugged.
Streamed tool calls and image input work on both surfaces, and service_tier is echoed back in usage so you can tell which queue served you.
# one endpoint, either SDKexport OPENAI_BASE_URL=https://api.partitionlabs.ai/v1export ANTHROPIC_BASE_URL=https://api.partitionlabs.aiclient = OpenAI() # unchanged stream = client.chat.completions.create( model="qwen3.8-27b", # one id per model messages=[{"role": "user", "content": "..."}], service_tier="flex", # queue order, not speed stream=True,)What flat rate actually means here
No token usage caps
Under contention, under-share tenants go first and over-share tenants queue. Nobody is rejected for using what they bought.
Cache writes are free
Prefix-cache reads cost almost nothing and writes cost nothing at all — a write is an ordinary prefill you already paid for. Efficient agent loops win.
Priority is queue order, never speed
Flex is the same model at the same tokens per second, last in line. Priority is a queue-jump under load. No stream is ever throttled.
And here is that queue, measured
Median time to first token per band over the last 24 hours, from served requests. On a quiet pool the bands converge — that is the system working, and we publish it either way.
No streaming traffic in the last 24 hours, so there is nothing honest to show here yet. The table fills in as the pool carries load.
Efficient intelligence for everyone.
We're an applied AI research and development company making efficient intelligence available to everyone.
Resources are finite, so we are efficient with them. That is the premise rather than a caveat: a card has twenty-four hours in a day and holds a measured number of always-on requests. We sell an honest share of it, and the research — training, inference, harnesses — is what gets more out of the same hardware.
One agent, running all month.
Metered, that bill grows with every token the agent reads, all month. On a subscription it is flat, from $10.00 a month, with the share written into the ledger above.