[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-fireworks-ember-1-kimi-k3-token-efficiency":3,"topics-all":38,"news-related-4c9f495a-bf44-4655-8b2b-29b4859a0af3":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"4c9f495a-bf44-4655-8b2b-29b4859a0af3","Ember-1:Kimi K3同质量,少四成token","Fireworks Research 基于 Kimi K3 重训出专用模型 Ember-1：官方实测七个基准加两家客户生产流量，推理长度缩短 35%-50% 而准确率不降；DeepSWE 1.1 得 75.2% 反超 K3 最高档，reasoning token 最多省 71.3%。现以研究预览上线。","推理模型的账单里，最贵的部分往往不是答案，而是思考。Fireworks 在官方博客里给了一个数字：像 Kimi K3 这样的推理模型，生成的 token 有时超过 90% 花在内部推理上，答案只占零头。单次请求贵一点还能忍，但在多轮 Agent 工作流里，每一轮都会把之前的推理重新喂回模型，上下文成本随轮数近似二次方增长——早期轮次的思考链会在后续每次调用里被重读、重新计费。\n\n## Ember-1：不调低思考强度，而是教会模型少想\n\n便宜的解法本来就有：把 K3 的 reasoning effort 调到低档。但 Fireworks 试过，低档位丢掉的质量太多。Ember-1 走另一条路——不动推理开关，而是重新训练模型本身。Fireworks Research 跑了 50 多个训练实验、200 多次评估，在数学、编码、指令遵循、对话、搜索、工具调用和软件工程等任务上训练 K3，目标只有一个：删掉不改变答案的推理，保留有价值的自我修正。官方口径是，七个公开基准加两家客户生产流量实测，K3 的推理长度可缩短 35%-50% 而准确率不降。Ember-1 是 Fireworks 自家模型系列第一款，现以研究预览形式上线 Serverless，两周窗口后视需求转正。\n\n## 数字长什么样\n\n基准表里最有信息量的是这组对比：Terminal Bench 2.1 上 Ember-1 拿 82.0%，反超 K3 默认最高档的 80.9%，同时省 51.9% 的 token；DeepSWE 1.1 上 75.2% 对 66.4%，同样反超；SWE-bench Verified 上 92.2% 对 93.2%，小输一个点但省 15.5% token。生产 A\u002FB 测试更直观：一家客户的编码工作流里，Ember-1 得分 0.753 对 K3 的 0.751，输出 token 从 49.3K 降到 29.9K，reasoning token 减少 71.3%，总 token 减少 39%。Fireworks 还先在自家内部编码流量上切换了模型，官方的说法很妙：开发者没注意到模型换了。\n\n## 小字部分才是重点\n\n独立分析（OrcaRouter 的评测博客）泼了几盆有用的冷水。第一，所有数字都是厂商自报，截至该文发出没有第三方复现。第二，「约 40%」是平均值，实际省幅从 τ-2 Bench Airline 的 5.9% 到 Terminal Bench 的 51.9% 不等——短的单轮调用基本感受不到节省。第三，Ember-1 未公布独立价卡，基准表省钱金额按 K3 公开价折算，而 OpenRouter 模型页挂出的价格与 K3 相同——省 token 就等于省钱。\n\n## 所以呢\n\nEmber-1 的模板值得记住：拿一个已经足够好的开放权重模型，只重训「思考长度」这一个轴，把省下的 token 变成差价。Fireworks 官方说 token 效率将是他们的主题，开放模型的未来是针对特定负载训练的专用模型。对在 Agent 循环里跑 K3 的团队：别只看榜单，拿自己的流量做影子测试，一两周后再决定要不要切换。毕竟压缩掉浪费的思考是白赚，压缩掉模型真正需要的步骤，代价会在几轮之后才以错误的工具调用形式浮现。\n\n参考：[Fireworks 官方公告](https:\u002F\u002Ffireworks.ai\u002Fblog\u002Fember-1) · [OrcaRouter 独立分析](https:\u002F\u002Fwww.orcarouter.ai\u002Fblog\u002Fember-1-release) · [OpenRouter 模型页](https:\u002F\u002Fopenrouter.ai\u002Ffireworks\u002Fember-1)","https:\u002F\u002Ffireworks.ai\u002Fblog\u002Fember-1","b53e6ed0-c594-4231-8fa3-27d83282f314",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":22,"name":23,"slug":23,"description":14,"color":14},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"9b28f7a6-4a7e-4340-80bc-7b1f3cd7f418","en","Ember-1: Kimi K3 Quality With 40% Fewer Tokens","Fireworks retrains Kimi K3 into Ember-1: 35-50% shorter reasoning at equal accuracy, 71.3% fewer reasoning tokens in production A\u002FB tests.","The most expensive part of a reasoning model's bill is often not the answer but the thinking. Fireworks gives a number in its official blog: reasoning models like Kimi K3 can spend more than 90% of their generated tokens on internal reasoning before writing anything a user sees. That is merely expensive on a single request, but it compounds in multi-turn agentic workloads — every turn replays prior reasoning back to the model, so context cost grows roughly quadratically with turn count. Early-turn thinking traces get re-read and re-billed on every subsequent call.\n\n## Ember-1: teach the model to think less, don't dial effort down\n\nThe cheap fix already existed: turn K3's reasoning effort down to low. Fireworks tried it; the low setting gave up too much quality. Ember-1 takes another path — leave the effort switch alone and retrain the model itself. Fireworks Research ran more than 50 training experiments and over 200 evaluations, training K3 across mathematics, coding, instruction following, conversation, search, tool use, and software engineering, with one goal: remove reasoning that does not change the answer while keeping self-correction that does. The official claim: across seven public benchmarks and two customers' production traffic, K3's reasoning could be shortened by 35–50% without sacrificing accuracy. Ember-1 is the first in Fireworks' own model series, live as a research preview on Fireworks Serverless with a two-week window, made permanent based on community demand.\n\n## What the numbers show\n\nThe benchmark sheet's most informative rows: on Terminal Bench 2.1, Ember-1 scores 82.0%, beating K3's default max of 80.9% while using 51.9% fewer tokens; on DeepSWE 1.1 it is 75.2% versus 66.4%, another outright win; on SWE-bench Verified it posts 92.2% against 93.2%, narrowly behind but saving 15.5% of tokens. The production A\u002FB test is more tangible: in one customer's coding workload, Ember-1 scored 0.753 versus K3's 0.751, with output tokens down from 49.3K to 29.9K — a 71.3% reduction in reasoning tokens and 39% in total tokens. Fireworks also switched its own internal coding traffic first; the official punchline is that developers never noticed the model had changed.\n\n## Read the fine print\n\nAn independent analysis (OrcaRouter's review blog) pours some useful cold water on the launch. First, every number is vendor-reported, with no third-party reproduction as of that writing. Second, the \"about 40%\" headline is an average: actual savings range from 5.9% on τ-2 Bench Airline to 51.9% on Terminal Bench — if your workload is short single-turn calls, you will barely feel it. Third, Ember-1 is a research preview with no independent rate card of its own; the dollar savings in the benchmark table are derived from K3's public pricing, and the OpenRouter model page lists the same price as K3 — meaning saved tokens equal saved money.\n\n## So what\n\nThe Ember-1 template is worth remembering: take an open-weight model that is already good, retrain exactly one axis — thinking length — and sell the token delta. Fireworks says token efficiency is becoming a theme for them, and that the future of open models is specialized models trained on specific workloads. For teams running K3 in agent loops, the pragmatic advice is to skip the leaderboard and shadow-test on your own traffic: send a share of real requests to both models and wait a week or two before switching anything. Compressing wasted deliberation is free profit; compressing a step the model actually needed shows up turns later as a wrong tool call.\n\nReference: [Fireworks announcement](https:\u002F\u002Ffireworks.ai\u002Fblog\u002Fember-1) · [OrcaRouter analysis](https:\u002F\u002Fwww.orcarouter.ai\u002Fblog\u002Fember-1-release) · [OpenRouter model page](https:\u002F\u002Fopenrouter.ai\u002Ffireworks\u002Fember-1)","fireworks-ember-1-kimi-k3-token-efficiency","2026-09-24T23:15:00Z","2026-09-24T23:10:13.007781Z","2026-09-24T23:10:13.007791Z",true,"agent",655,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"a574d32a-c8f7-4d46-ac95-b76406f36b3b","OpenAI 把 GPT-6 拆成两档:Sol\u002FLuna 同款训练法,API 价格砍一半","openai-gpt-6-sol-luna-api-pricing","2026-09-23T00:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"b667e52f-ec7d-4ca4-8d9e-1db81e1a5616","DeepSeek论文:890字节KV缓存的三层架构账","deepseek-v41-flash-kv-cache-paper","2026-09-18T15:10:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"4c4a2a9e-f69b-4985-bd42-97ab2ef4e2ac","Spark-X2.5-4B 开源:4B 跑 1M 上下文,22 项基准打 9B 级 Qwen3.5","spark-x2-5-4b-apache-open-source","2026-09-16T01:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"ee62535f-b897-437b-8674-02801632dadb","DeepSeek V4.1 非对称架构首发:读题 8B 答题 16B,KV 缓存砍到初代的 1\u002F437","deepseek-v4-1-flash-ced-kv-cache","2026-09-10T15:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"ee700812-e4e5-4fa5-8687-6a0d7e5c7f78","Mercury 2.5 发布：扩散 LLM 跑出 1107 tokens\u002F秒，智能较上代提升 40%","mercury-2-5-diffusion-llm","2026-09-09T17:05:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"491f4904-c854-4925-b3e3-e34b8afd5e50","KDA+MLA 混合栈下沉到 1.3B 激活:Ling-3.0-tiny 把 MoE 端侧化,INT4 跑出 115 tok\u002Fs","ling-3-tiny-kda-mla-edge-deployment","2026-08-18T00:00:00+00:00"]