Ember model releases

Ember-1 is the first model from Fireworks Research, released 2026-09-23 as a Research Preview on Fireworks' serverless platform (a two-week window, made permanent based on demand). It is a specialized reasoning model post-trained on Moonshot's open-weight Kimi K3 (via 50+ training experiments and new training algorithms, not a simple fine-tune), designed to deliver roughly Kimi K3 quality while emitting ~35-50% shorter reasoning traces (about 40% fewer output tokens in live A/B tests) — targeting coding, knowledge work, and agentic workflows where reasoning cost and latency matter. Text input and output; 1,048,576-token context window. Pricing $3 / 1M input ($0.30 cached), $15 / 1M output (matching Kimi K3 on Fireworks). Reported benchmarks vs Kimi K3: Terminal-Bench 2.1 82.0 (vs 80.9), DeepSWE 1.1 75.2 (vs 66.4), tau2-Bench Airline 66 (vs 64), SWE-bench Verified 92.2 (vs 93.2). Ember-1's own weights are not published (hosted preview); its Kimi K3 base is open-weight.