What Is Mercury 3?Mercury 3 是什么？
Mercury 3 is the diffusion language model (dLLM) from Inception (Redwood City), released October 19, 2026. Unlike autoregressive models that generate one token at a time, a diffusion LLM drafts many tokens in parallel and refines them over a few passes — Inception reports 1,068 tokens per second on NVIDIA GPUs, a 1M-token context window (up from 260K on Mercury 2.5), up to 131,072 output tokens, and more than 40x the total parameters of Mercury 1 at comparable speed. Vendor benchmarks claim quality on par with or better than GPT-6 Luna and Claude Haiku 5.5 while running 4–8x faster — vendor-reported, not independently verified.Mercury 3 是 Inception（Redwood City）2026 年 10 月 19 日发布的扩散式语言模型（dLLM）。和一次生成一个 token 的自回归模型不同，扩散式 LLM 并行起草大量 token、再经几轮精炼——Inception 报告在 NVIDIA GPU 上跑出 1068 tokens/s、100 万上下文（Mercury 2.5 的 260K 上调而来）、最长 131,072 输出 token，总参数是 Mercury 1 的 40 倍+而速度相当。厂商基准称质量打平或超越 GPT-6 Luna 与 Claude Haiku 5.5、快 4–8 倍——厂商自报数据，未经独立验证。
Pricing: $0.10 input / $0.35 output per million tokens at launch (50% off the $0.20/$0.70 list prices — no published end date for the discount), cached input 90% off, and 100M free tokens for new API accounts. It is proprietary with no open weights — API only. Prices per the official site.定价：首发 $0.10 输入/$0.35 输出每百万 token（原价 $0.20/$0.70 的 5 折——官方未公布折扣截止日），缓存输入 90% off，新 API 账号送 100M 免费 tokens。专有模型、无开权重——只能调 API。价格以官网为准。
Mercury 3 is the strongest evidence yet that diffusion LLMs can scale: 1,068 tokens/s at more than 40x the size of Mercury 1 attacks the exact bottleneck that kills agent UX — every agent step waits on the model, and parallel decoding collapses that wait. The launch pricing ($0.10/$0.35 per MTok, cached input 90% off, 100M free tokens) makes it nearly free to evaluate, and the OpenAI-compatible API plus OpenRouter/Baseten/Vercel distribution means near-zero integration cost. The honest tradeoffs: all quality claims are vendor-reported (8 of 12 vendor-run benchmarks led, but Haiku 5.5 still wins DeepSWE v1 and MRCR v2), there are no open weights at all, and the launch discount has no published end date — budget at list prices ($0.20/$0.70) if your planning horizon is long. Verdict: if your bottleneck is agent latency, this is the cheapest experiment in the market right now — spin it up on OpenRouter with the free tokens before the discount math changes.
Mercury 3 是扩散式 LLM 能 scale 的最强证据：1068 tokens/s、体量是 Mercury 1 的 40 倍+，打的正是杀死 Agent 体验的瓶颈——Agent 每一步都在等模型，并行解码把等待压扁。首发价（$0.10/$0.35 每百万 token、缓存输入 90% off、100M 免费 tokens）让试用几乎零成本；OpenAI 兼容 API 加上 OpenRouter/Baseten/Vercel 全渠道，集成成本也接近零。真实代价：质量数据全是厂商自报（12 项厂商基准赢 8 项，但 DeepSWE v1 和 MRCR v2 仍是 Haiku 5.5 拿下）；完全没有开权重；首发折扣没有公布截止日——规划周期长的按原价（$0.20/$0.70）做预算。结论：如果你的瓶颈是 Agent 延迟，这是现在市场上最便宜的实验——趁折扣数学还没变，先在 OpenRouter 上用免费额度跑起来。
— 中国吧 AI Tools Hub Editorial Team
Who Should Use Mercury 3?谁适合使用Mercury 3？
✓ Good Fit For适合以下场景
- Agent builders — every model call's latency compounds across stepsAgent 开发者——每一步的模型延迟都会累积
- Long-document work needing the full 1M context at a low price长文档场景：要 100 万上下文，又要低价
- Teams already on OpenAI-compatible stacks — near-zero integration cost已在 OpenAI 兼容技术栈上的团队——集成成本接近零
✕ Not Ideal For不适合以下场景
- Anyone needing open weights or self-hosting — none exist需要开权重/自托管的人——没有
- Buyers who need independently verified benchmarks before committing提交前需要独立验证基准的买家——先等等
- Long-horizon budget planning on the launch discount — no end date published按首发折扣做长期预算——官方没公布截止日
Getting Started with Mercury 3Mercury 3 快速开始
1. Grab the free tokens: create an API key at platform.inceptionlabs.ai — 100M free tokens for new accounts领免费额度：在 platform.inceptionlabs.ai 创建 API key——新账号 100M 免费 tokens2. Point your OpenAI-compatible client at it: model string mercury-3 (Inception API) or inception/mercury-3 (OpenRouter)OpenAI 兼容客户端直接指过去：模型名 mercury-3（Inception API）或 inception/mercury-3（OpenRouter）3. Tune reasoning_effort across 4 levels; test on your real agent loop before budgeting at list prices ($0.20/$0.70 per MTok — prices per official site)reasoning_effort 在 4 档里调；在你真实 Agent 链路上实测，再按原价（$0.20/$0.70 每百万 token，价格以官网为准）做预算Key Features核心功能
-
1,068 Tokens/s Parallel Decoding1068 tokens/s 并行解码 — Diffusion decoding drafts many tokens in parallel instead of one at a time — 1,068 tok/s on NVIDIA GPUs, 4–8x faster than the speed-tier models Inception benchmarks against.扩散解码并行起草大量 token，而不是一次一个——NVIDIA GPU 上 1068 tokens/s，比 Inception 对标的速度档模型快 4–8 倍。
-
1M-Token Context100 万上下文 — A 1M-token context window (up from 260K on Mercury 2.5) with up to 131,072 output tokens — whole codebases and long documents in one pass.100 万 token 上下文（Mercury 2.5 的 260K 上调而来），最长 131,072 输出 token——整个代码库、超长文档一次吃下。
-
OpenAI-Compatible APIOpenAI 兼容 API — Drop-in replacement for OpenAI-style endpoints: tool calling, structured outputs and a reasoning_effort parameter with four levels. Served on the Inception API (mercury-3), OpenRouter (inception/mercury-3), Baseten and Vercel.OpenAI 风格接口的直接替代：工具调用、结构化输出、4 档 reasoning_effort 参数。上线 Inception API（mercury-3）、OpenRouter（inception/mercury-3）、Baseten、Vercel。
-
Launch Pricing: 50% Off首发 5 折定价 — $0.10 input / $0.35 output per MTok at launch (list $0.20/$0.70), cached input 90% off, 100M free tokens for new accounts. Prices per official site.首发 $0.10 输入/$0.35 输出每百万 token（原价 $0.20/$0.70），缓存输入 90% off，新账号 100M 免费 tokens。价格以官网为准。
-
40x+ the Size of Mercury 1Mercury 1 的 40 倍+体量 — More than 40x Mercury 1's total parameters at comparable speed — Inception's case that diffusion scaling is limited by compute and data, not the algorithm.总参数是 Mercury 1 的 40 倍+，速度相当——Inception 论证扩散式扩展的瓶颈是算力和数据，不是算法。
Pros & Cons优缺点
✓ Good Fit For适合以下场景
- Fastest-feeling API tier for agent loops: 1,068 tok/sAgent 链路里体感最快的 API 档：1068 tokens/s
- 1M context + launch pricing at $0.10/$0.35 per MTok100 万上下文 + 首发价 $0.10/$0.35 每百万 token
- Every major channel: Inception API, OpenRouter, Baseten, Vercel全渠道：Inception API、OpenRouter、Baseten、Vercel
- Near-zero integration cost via OpenAI compatibilityOpenAI 兼容，集成成本接近零
✕ Not Ideal For不适合以下场景
- Proprietary — no open weights, no self-hosting, period专有——无开权重，不能自托管，没得商量
- Quality claims are vendor-reported, not independently verified质量数据厂商自报，未经独立验证
- Launch discount end date unpublished — budget at list prices for long horizons首发折扣无截止日——长期规划按原价做预算
Use Cases应用场景
Popular ways people use Mercury 3.人们使用Mercury 3的常见场景：
🤖 Browser & Computer-Use Agents浏览器与 Computer-Use Agent
Every agent step waits on the model — 1,068 tok/s collapses per-step latency so ten-step tasks stop feeling eternal.Agent 每一步都在等模型——1068 tokens/s 把单步延迟压扁，十步任务不再像等到天荒地老。
🔎 Real-Time Search Answers实时搜索问答
Sub-second generation (0.9s median in Inception's Octen AI example) for live search products where speed is the feature.亚秒级生成（Inception 的 Octen AI 示例中位数 0.9 秒）——速度本身就是功能的实时搜索产品。
📚 Long Documents at Low Cost低成本长文档处理
The full 1M context at launch prices ($0.10/$0.35 per MTok; prices per official site) — long-document work without the flagship-tier bill.100 万上下文配首发价（$0.10/$0.35 每百万 token；价格以官网为准）——长文档任务不用付旗舰档账单。
💬 Interactive Chat & Education交互式聊天与教育
Chat products and tutoring that need fast, fluent responses — Mercury Chat for consumers, API for builders.需要又快又流畅回复的聊天产品和教学场景——消费者用 Mercury Chat，开发者调 API。
Similar AI Tools相似 AI 工具
If Mercury 3 doesn't fit your needs, here are other popular AI tools you might consider — for a comparable API-first flagship see our full review: Mistral Large 4.如果Mercury 3不符合你的需求，可以看看这些热门 AI 工具——同类 API 优先旗舰可参考本站完整评测：Mistral Large 4。
Frequently Asked Questions常见问题
mercury-3), OpenRouter (inception/mercury-3), Baseten and Vercel, with an OpenAI-compatible API, tool calling, structured outputs and a reasoning_effort parameter. Mercury Chat is the consumer chat app for trying it.Mercury 3 仅 API（无开权重）。上线 Inception API 平台（模型名 mercury-3）、OpenRouter（inception/mercury-3）、Baseten 和 Vercel；OpenAI 兼容 API，支持工具调用、结构化输出和 reasoning_effort 参数。消费者可以用 Mercury Chat 直接体验。