Fireworks launches Ember-1: Kimi K3-based model cuts reasoning tokens by ~40%
sophiamyang · x · 2026-09-24
Fireworks Research released Ember-1, a specialized model built on Kimi K3 that curbs over-reasoning: 40% fewer tokens with comparable quality.
- Method: 50+ training runs and 200+ evals on Fireworks serverless training, using new algorithms to cut wasted reasoning traces.
- Results: Matches k3-max quality on 7 public benchmarks; best cost-quality tradeoff vs GPT-6 Astra and Opus 5 on Bedside Bench; one customer is moving over entirely.
- Specs & pricing: 2.78T-param MoE, 1040k context, function calling and image input supported; serverless at $3.00/M input, $0.30/M cached input, $15.00/M output, plus on-demand GPU deployment.
More from Infra
- UK Sovereign AI invests in Fractile to build next-gen frontier AI chips — charlieharris01 · 2026-11-11
- What happened when I moved vector retrieval into Snowflake SQL instead of orchestrating RAG outside — dima806_dima · 2026-11-11
- AI chip startup Fractile raises $720M Series C for DRAM-on-processor design — Tom_Westgarth15 · 2026-11-11
- Mobile AI Factory: agentic pipeline builds on-device SLMs that beat 5-shot frontier models — kalyan_kpl · 2026-11-11
- Vivo's Starlink direct-to-cell service hits 10M users in two weeks across Brazil — XFreeze · 2026-11-11
- Nan Ya Plastics to spend NT$11.53B on fifth glass fabric plant on AI server demand — zephyr_z9 · 2026-11-11