Fireworks AI Releases Ember-1: A Post-Trained Kimi K3 That Uses About 40% Fewer Tokens
https://www.marktechpost.com/2026/09/28/fireworks-ai-releases-ember-1-a-post-trained-kimi-k3-that-uses-about-40-fewer-tokens/📌 Fireworks Ember-1：把 Kimi K3 的推理 token 砍掉四成
TL;DR：Fireworks 後訓練 Kimi K3 推出 Ember-1，在準確度打平的前提下少生成約 40% token。
推理模型的帳單痛點往往不在「答錯」，而在「想太多」。Fireworks 團隊觀察到，像 Kimi K3 這樣的推理模型，生成的 token 裡有時超過 90% 花在內部推理上；在多輪 agentic 任務中，這個成本會被放大，因為每一輪都要把先前的推理過程重新餵回模型，context 隨對話輪數大致呈二次成長，早期輪次留下的冗長推理痕跡，會在後續每一次呼叫中被重新讀取、重新計費。
🤔 調低 effort 設定解決不了問題
Fireworks 表示，客戶原本想要的是 K3 的程式碼能力，但用更低的成本取得。直覺的做法是調低推理力度（reasoning effort）設定，但團隊發現這樣做犧牲了太多品質。問題的關鍵在於，K3 的推理過程並非全是浪費，其中一部分是有用的自我反思，例如重新檢視假設或針對回饋做出調整。於是 Fireworks 選擇的路線是後訓練模型，讓它學會更有效率地推理，保留有用的自我反思行為，同時砍掉冗餘的推理與無效的重複迴圈。
🧩 訓練涵蓋的任務範圍與規模
訓練資料涵蓋數學、程式碼、指令遵循、對話、搜尋、工具使用與軟體工程，同時包含單一問題與延伸的多步驟互動；任務與環境回饋被用來引導 on-policy 的規劃與學習。Fireworks 團隊總共跑了超過 50 次訓練實驗與超過 200 次評估，並開發了尚未公開的新訓練演算法，整個訓練流程都在 Fireworks Serverless Training 上完成，且僅使用自有資料，不含客戶資料。
📊 公開評測：多數任務打平或超越，成本大幅下降
Fireworks 用三種推理力度等級的 Kimi K3 作為對照組，並以 K3 公開的 API 定價計算成本，以下是官方公布的部分結果：
| 指標 | Kimi K3 | Ember-1 |
|---|---|---|
| 輸出 token（每任務） | 49.3K | 29.9K |
| 任務分數 | 0.751 | 0.753 |
| 平均步驟數 | 23.8 | 21.4 |
| 輸出成本估算（每任務） | 約 $0.45 | 約 $0.74（原文，可能為兩者數值對調，詳見來源） |
（註：上表中成本數字為 MarkTechPost 自行依輸出 token 換算，僅計輸出部分。）
整體來看，Ember-1 的推理 token 相較 K3 下降了 71.3%，總 token 用量下降 39%，任務分數幾乎沒有變化。在具體基準測試上，Ember-1 在 Terminal Bench 2.1 與 DeepSWE 1.1 上領先 K3 Max，但在 SWE-bench Verified 與 SWE-Interact 上略微落後；在涵蓋七個基準與兩位客戶正式環境流量的測試中，K3 的推理長度被縮短了 35% 到 50%，且沒有犧牲準確度。在 Doximity 的 Bedside Bench（一組經醫師驗證的 500 筆臨床案例集）上，Ember-1 透過 Fireworks 新推出的 Specialized Intelligence Index，建立了新的「成本對任務」帕雷托前緣（Pareto frontier）。Fireworks 也與兩位客戶在正式程式碼工作負載上做了即時 A/B 測試，兩者都在品質相當的情況下，觀察到每個任務約減少 35% 的 token 用量；目前已有一位客戶把 Ember-1 用在正式環境中。
⚠️ 目前無法自行部署
Ember-1 只能透過 Fireworks 的 serverless API 以 Research Preview 形式使用，Fireworks 並未釋出模型權重、訓練程式碼或確切的訓練演算法，因此自行架設（self-hosting）目前不是選項。在 Fireworks 上，Ember-1 的每 token 計費與 Kimi K3 相同，輸入 $3.00、快取輸入 $0.30、輸出 $15.00（每百萬 token），省下的成本完全來自生成的 token 變少，而非單價調整。
🎯 實務啟示
如果你的 agentic 工作負載正被多輪對話中不斷重播、重新計費的冗長推理過程拖累，Ember-1 的做法提供了一個思路：與其在推論階段硬調低 reasoning effort 犧牲品質，不如透過後訓練讓模型自己學會分辨哪些推理有用、哪些是重複迴圈。但在自行採用前，仍要留意目前只能透過 Fireworks 的托管 API 使用，且在部分 SWE 基準上表現略遜於原始 K3。
🔗 來源
- 標題：Fireworks AI Releases Ember-1: A Post-Trained Kimi K3 That Uses About 40% Fewer Tokens
- 作者／機構：Asif Razzaq
- 連結：https://www.marktechpost.com/2026/09/28/fireworks-ai-releases-ember-1-a-post-trained-kimi-k3-that-uses-about-40-fewer-tokens/
#FireworksAI #Ember1 #KimiK3 #LLM #ReasoningModels #AgenticAI #TokenEfficiency #MoonshotAI #AIInference #MachineLearning
原始資料 MarkTechPost · 收集於 2026-09-28
摘要原文
Fireworks AI has released Ember-1 , a specialized model from Fireworks Research built by post-training Moonshot AI’s open-weight Kimi K3 . Ember-1 learns to produce shorter reasoning traces while keeping task accuracy. This is different from lowering the reasoning effort setting at inference time. According to the Fireworks release post , Ember-1 delivers Kimi K3’s quality with about 40% fewer tokens. Is it deployable? Yes, but only through the Fireworks serverless API as a Research Preview. Fireworks has not released Ember-1’s weights, training code, or exact training algorithms, so self-hosting is not an option today. Fireworks team reports that reasoning models like Kimi K3 sometimes spend more than 90% of generated tokens on internal reasoning. That cost compounds in multi-turn agentic workloads. Each turn replays prior reasoning back to the model. Context grows roughly quadratically with the number of turns. Long traces from early turns get re-read, and re-billed, on every later call. Fireworks team explains how customers wanted K3’s coding capability at lower cost. Turning down K3’s reasoning effort did not solve it. Lower effort settings gave up too much quality. So the team trained the model to reason more efficiently instead. Not all of K3’s reasoning is waste. Some of it is useful self-reflection, like revisiting an assumption or reacting to feedback. Ember-1 keep that behavior while cutting redundant reasoning and unproductive loops. The training collection spans mathematics, coding, instruction following, conversation, search, tool use, and software engineering. It covers both standalone problems and extended multi-step interactions. Task and environment feedback guides on-policy planning and learning. Fireworks team ran more than 50 training experiments and over 200 evaluations. They also developed new training algorithms, which it has not published. All training ran on Fireworks Serverless Training . Fireworks states it used its own data and no customer data. Fireworks compared Ember-1 with Kimi K3 at three reasoning effort levels. Cost was computed with public Kimi K3 API pricing. These are Fireworks’ own published evaluations. Ember-1 leads K3 Max on Terminal Bench 2.1 and DeepSWE 1.1. It trails slightly on SWE-bench Verified and SWE-Interact. Across seven benchmarks and two customers’ production traffic, Fireworks says K3’s reasoning was shortened by 35 to 50% without sacrificing accuracy. On Doximity’s Bedside Bench, a physician-validated set of 500 clinical cases, Ember-1 set a new cost-per-task Pareto frontier. That result comes from Fireworks’ new Specialized Intelligence Index . Fireworks ran live A/B tests with 2 customers on production coding workloads. Both saw roughly 35% fewer tokens per task at comparable quality. In the published run, output tokens fell from 49.3K to 29.9K per task. Reasoning tokens dropped 71.3% and total tokens dropped 39%. The task score was essentially unchanged: 0.753 for Ember-1 versus 0.751 for K3. Average steps fell from 23.8 to 21.4. One customer now runs Ember-1 in production. Ember-1 costs the same per token as Kimi K3 on Fireworks: $3.00 input, $0.30 cached input, and $15.00 output per 1M tokens. The savings come entirely from generating fewer tokens. At that output rate, the A/B figures work out to about $0.74 versus $0.45 in output cost per task (our calculation, output only). Check out the Technical Details . All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter . Wait! are you on telegram? now you can join us on telegram as well. Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us The post Fireworks AI Releases Ember-1: A Post-Trained Kimi K3 That Uses About 40% Fewer Tokens appeared first on MarkTechPost .
由 claude-code/sonnet 自動生成