Ember-1

Fireworks announced Ember-1 on September 23, 2026, as a specialized model post-trained from Moonshot AI's Kimi K3 open weights, training reasoning length as a behavior rather than leaving it as a fixed cost of model quality. The objective was narrower than building a new foundation model: teach K3 to use shorter reasoning traces without abandoning useful analysis. Fireworks reports completing more than 50 training experiments and over 200 evaluations across mathematics, coding, instruction following, conversation, search, tool use, and software engineering, using task and environment feedback for on-policy learning. However, the evidence is primarily vendor-run, and Ember-1 is API-only with no released weights or reproducible training pipeline, leaving outside researchers unable to independently inspect or reproduce the result.

Fireworks AI released Ember-1, a specialized model from Fireworks Research that post-trains Moonshot AI's open-weight Kimi K3 to produce shorter reasoning traces while preserving task accuracy. According to Fireworks' release, Ember-1 delivers K3's quality with roughly 40% fewer tokens, addressing customer complaints that reasoning models like K3 sometimes spend over 90% of tokens on internal reasoning. Ember-1 is deployable only via the Fireworks serverless API as a Research Preview, with no released weights, training code, or exact training algorithm, so self-hosting is not currently possible. Fireworks reports that lowering K3's inference-time reasoning effort lost too much quality, so the team instead trained Ember-1 to cut redundant loops while keeping useful self-reflection, spanning math, coding, instruction following, and conversation workloads.

Fireworks AI released Ember-1 on September 23, 2026, a specialized model built by post-training Moonshot AI's 2.78-trillion-parameter Kimi K3 base weights. The post-training trims verbose reasoning traces, letting Ember-1 match K3 quality while using roughly 40% fewer tokens. List pricing sits at $3.00 input, $0.30 cached input, and $15.00 output per million tokens. Ember-1 targets the "double-billing trap" in multi-turn agent workflows, where reasoning tokens billed as output on turn one are re-sent as input on subsequent turns, compounding costs. The piece frames Ember-1 as Fireworks's playbook for curbing token inflation in coding-agent deployments. Note: the article carries promotional directory CTAs that do not affect the technical claims.

Fireworks announced Ember-1 on September 23, 2026, as a specialized model built by post-training Kimi K3 to produce shorter reasoning traces while preserving task quality. The company reports roughly 40 percent fewer tokens across its evaluations and about 35 percent fewer tokens per task in live customer A/B tests. Hosted on Fireworks Serverless at $3 per million input tokens and $15 per million output, Ember-1 ships as a time-limited research preview with two-week serverless access and no downloadable weights. All published comparisons are Fireworks-run, so developers should measure cost and failure rates against their own agentic workloads before swapping out K3.

Fireworks AI released Ember-1 on September 23, 2026, as the first model from Fireworks Research. Rather than a new base model, Ember-1 is Moonshot AI's open-weight Kimi K3 post-trained on Fireworks' training service to reason in fewer tokens. Fireworks reports it matches K3's quality with 35 to 50 percent shorter reasoning across seven benchmarks. Pricing stays identical to K3 on Fireworks at $3 per million input tokens, $0.30 cached, and $15 per million output, so any bill reduction comes purely from shorter outputs. Ember-1 is a research preview with a two-week serverless access window, no downloadable weights, and Fireworks-run benchmarks without independent replication.

Fireworks AI introduced Ember-1, a model further trained on Moonshot AI's Kimi K3 to address excessive reasoning-token consumption in cutting-edge models. Ember-1 is the first release in Fireworks's initiative to build developer-specialized models, targeting the problem that advanced models devote roughly 90% of output to thinking, inflating agent costs. According to Fireworks AI, Ember-1 retains Kimi K3 performance while cutting output tokens by approximately 35%. On the Bedside Bench medical task benchmark, Ember-1 matched Kimi K3 scores at lower per-task cost. The base Kimi K3, released July 2026, had surpassed GPT-5.6 Sol and Claude Fable 5 on some tests but ran slower and pricier, especially on the AA-Briefcase agent benchmark.