Ember-1: benchmark results

Fireworks AI's Ember-1 (September 2026), Kimi K3 post-trained by Fireworks Research to reason in about 40% fewer tokens; a Research Preview on the Fireworks serverless API with no released weights. Provider: Fireworks AI. Released 2026-09-23. Access: API.

Aggregate Rating 1800 ± 16, rank #90 of 10556 rated models, from 29 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
LLM Stats (DeepSWE 1.1)75.2Score (%)83.3
LLM Stats Score41.94LLM Stats Score (conservative rating)83.1
Vals AI Vibe Code Bench83.87Accuracy (%)82
DuelLab GameBench 2 - Reasoning Variants63GameBench2 normalized score (0–100)73.5
Gert Labs Rankings53.68GScore (%)73.5
Vals AI Legal Research Bench44.71Accuracy (%)73.4
Vals AI Harvey Legal Agent Bench9.17Accuracy (%)71.9
Tinybird AI SQL Benchmark - Success Rate100Questions answered with a valid query within 3 retries (%)71.8
ObviousBench97.22Answer pass³ (%)71.4
Tinybird AI SQL Benchmark - Exactness51.73Result exactness vs human reference queries (0-100)70.6
Vals AI Excel Modeling63.37Accuracy (%)64
Vals AI Tax Agent Bench31.08Accuracy (%)62.4

Interactive version: theaggregate.ai/model?slug=ember-1 · How It Works · Data refreshed daily, snapshot 2026-11-17.