Fireworks Ember-1 Cuts Reasoning Token Use 40% on Kimi K3 Base at $3/$15 Per Million
Fireworks Research released Ember-1 on September 23, a reasoning model built on Kimi K3, the 2.8 trillion parameter open-weight model from Moonshot AI. The defining claim: Ember-1 produces reasoning traces that are 40% shorter than Kimi K3’s baseline while maintaining comparable quality across Fireworks’ benchmark suite.
Specs
- Base model: Kimi K3 (Moonshot AI, 2.8T MoE open-weight)
- Pricing: $3 input / $15 output per 1M tokens
- Context: 1M tokens
- Inputs: Text and images
- Released: September 23, 2026
Pricing matches the base Kimi K3 rate. The token reduction is the efficiency gain — shorter reasoning traces mean fewer billed tokens per query at the same cost per million.
The Efficiency Case
Kimi K3 scores 93.4% on SWE-bench Verified, placing it third on that leaderboard behind Claude Opus 5 and DeepSeek V4 Pro. Its reasoning capability is frontier-grade. Ember-1’s premise is that most production tasks do not require the full depth of Kimi K3’s reasoning traces — and that a model fine-tuned to find shorter paths can deliver similar output at lower token cost.
The 40% figure comes from Fireworks’ internal evaluations on coding and knowledge work tasks. No independent benchmark runs have been published. Fireworks describes the quality gap as “comparable” without specifying a threshold.
Why Agentic Workflows
Token overhead in reasoning models compounds in agentic loops. A coding agent making 20 sequential calls to a model that generates 2,000 thinking tokens per call accumulates 40,000 reasoning tokens per task. At 40% reduction, that drops to 24,000. At $15/M output, the saved 16,000 reasoning tokens cut $0.24 per task — meaningful across thousands of daily agent runs.
Ember-1 is served via OpenRouter through a single provider. It is available for immediate use via OpenRouter’s API.
Context
At least two other labs productized Kimi K3 before Ember-1: Abacus.AI fine-tuned the same model in August into an agentic specialist that exceeded Kimi K3 on four separate benchmarks, and Cognition AI released SWE-2 on the same base in September. Moonshot AI released Kimi K3 under an open license in July, enabling downstream specialization. Ember-1 targets the opposite end of the trade-off: fewer tokens, same quality, lower cost per run rather than higher benchmark ceiling.