Applies to the next response. Responses use the engine’s remaining context.
History and preferences are saved for this browser. Export conversations to keep a copy.
Model
Keep keys private. Each secret is shown once.
You won’t be able to view it again.
OpenAI-compatible, streaming supported. API calls don’t appear in chat history.
SEINARI · RTX 5090
Choose Strict for near-original behavior of PrismML's Ternary Bonsai 2 27B or Fast for lossy speculative generation. Both have a separately tuned batch64 configuration.
Speed in tokens/s. Native is decode-only; HTTP includes prompt processing. Earlier-card results and tiny screens are labeled. Different workloads are not a paired comparison. “—” means unmeasured.
Includes controls, earlier iterations, profiling costs, accuracy regressions and rejected schedules. Values retain their original units and measurement scope.
Full retained reports from this investigation, including methodology, negative results, numerical checks and profiling explanations. Deployment statements inside earlier reports describe that stage of the investigation.
Run a short live probe on the same 5090. It shares the chat queue. Results are ungraded, and timing includes prefill and HTTP but excludes model loading. A single chat request does not exercise a full batch.