Surogate Rune 26B-A4B goes serverless on Fireworks at $0.042 per million input tokens

Last updated: October 29, 2026

Source · Model card: Surogate Rune 26B-A4B v3

publicado (OK do Fabricio 30/10 BRT).

Fireworks announced on 29 October 2026 (20:05 BRT) that Rune is live on its serverless API. The model page names it Rune 26B-A4B, points at the Hugging Face repo surogate/rune-26b-a4b-GGUF and lists 26.5B parameters (MoE), 262K context and text input only.

Price. $0.042 per million input tokens, $0.008 cached, output free. On-demand deployments on dedicated GPUs are also offered. The page says fine-tuning and image input are not supported.

Vendor numbers (UnverifiedClaim). Fireworks says a median latency under 150 ms measured server-side and up to 256 questions per pass. The Fireworks page marks the model “Calibrated: No”; the author fits a temperature of 2 on its own, and the Decision Index 0.4.1 measures ECE 0.12 at the default.

Boards (30 Oct). Unchanged: Decision Index 0.4.1 Full 56.42 (#70 of 247); JevBench Intelligence 56.1, calibration 90.9. Not on Artificial Analysis.

Score. The Surogate Rune page keeps its scores (5.8 / 8.1 / 11) for now; the Fireworks listing is a maturity argument we have flagged for review, not applied.