What is the best local AI?什么才是最好的本地 AI?

Is DGX Spark the only answer?DGX Spark 是唯一的答案吗?

$3,999 a year ago.一年前 $3,999,$6,999 today.今天 $6,999,Tomorrow?明天呢?

The new 64GB model costs more than the old 128GB.新出的 64GB,
比当初的 128GB 还贵。

"1,000 TOPS." Sparse. On paper."1000 TOPS",
是纸面上的稀疏算力。

And we were told to just accept it.而我们,只能默默接受。

Not anymore.不必再这样了。

We're breaking the monopoly. Win or lose.我们要打破垄断。
无论成败。

Freedom for personal local AI.为了个人本地 AI 的自由。

MaixStation

The Best Local AI. At Every Scale.本地 AI 最佳之选,任意规模。

From 27B to 552B Models.
From 3" Module to Cluster.
Beats DGX Spark.
从 27B 到 552B 大模型。
从 3 寸核心板,到迷你机架集群。
超越 DGX Spark。

128GB LPDDR5X273 GB/s40 W11×11×3 cm
MaixStation held in one hand, front panel facing out
MaixStation · 128GBFIG. 01
Tiny

Half a Mac mini.
A third of a DGX Spark.
Mac mini 的一半。
DGX Spark 的三分之一。

MaixStation363 cm³
Mac mini806 cm³
DGX Spark1,125 cm³

363 cm³. Small enough for your pocket.363 cm³,小到可以塞进口袋。

MaixStation, Mac mini and DGX Spark side by side
MaixStation · Mac mini · DGX SparkFIG. 02
Front view
Front正面110 × 30 mm
Side view
Side侧面110 × 30 mm
Rear view
Rear背面110 × 30 mm
Top view
Top顶面110 × 110 mm
Bottom view with intake grille
Bottom底面110 × 110 mm
Powerful

Beats DGX Spark.超越 DGX Spark。

AX8860P · 256 TOPS FP4128GB LPDDR5X · 273 GB/s

Qwen3.8‑27B DENSE · FP4

Prefill · tok/s
DGX Spark
0
MaixStation
0–0
Decode · tok/s
DGX Spark
0
MaixStation
0–0

Qwen3.6‑35B‑A3B MOE · FP4

Prefill · tok/s
DGX Spark
0
MaixStation
0–0
Decode · tok/s
DGX Spark
0
MaixStation
0–0
FP4 · MTP disabled · AX8860P is still being tuned. Bars show today's low to high; even the low end beats DGX Spark.FP4 · 未启用 MTP · AX8860P 仍在优化调试中,柱条为当前区间,最保守的性能也优于 DGX Spark。

256 TOPS beats "1,000 TOPS" in prefill?256 TOPS 为什么能赢"1000 TOPS"?

1,000 is sparse FP4, which real models can't use. Dense is 500 on paper, and on real models Spark puts only a small slice of it to work. MaixStation turns far more of its 256 into actual speed.1000 是稀疏 FP4 算力,真实模型用不上。稠密算力纸面也只有 500,跑真实模型时 Spark 只用上其中一小部分。MaixStation 能把 256 中大得多的比例变成实际速度。

Same 273 GB/s. Why is decode faster?同样 273 GB/s,为什么解码更快?

Bandwidth sets the ceiling. Architecture decides how close you get. Every token reads the whole model once, and MaixStation keeps more of those 273 GB/s busy.带宽决定上限,架构决定离上限有多近。每生成一个 token 都要把模型读一遍,MaixStation 把这 273 GB/s 用得更满。

tok/s ≈ effective bandwidth ÷ model size

I've seen Spark do 50+ tok/s on 27B.我见过 Spark 跑 27B 有几十 tok/s?

Those runs use MTP or speculative decoding. This chart turns MTP off on both sides for a like‑for‑like baseline. Turn it on, and MaixStation gets a similar speed‑up.那些数字开了 MTP 或投机解码。本图两边都关闭 MTP,用标准速率对比。开启 MTP 后,MaixStation 同样获得相近的加速比。

Efficient

40 watts.
A power bank is enough.
40 瓦。
充电宝就够。

MaixStation · typical load
40W
USB‑C PD. Any 65W bank.USB‑C PD。任何 65W 充电宝。
DGX Spark · typical load
120W
Dedicated brick. Wall outlet.专用电源砖。必须插墙。
Portable

On the desk.
In the pocket.
放在桌上。
装进口袋。

MaixStation sliding into a trouser pocket
MaixStation · in the pocket装进口袋FIG. 05
MaixStation front panel: power button, power USB-C, USB 3.0, mode switch, 10G SFP+
  1. Power电源键
  2. Power USB‑C供电 USB‑C
  3. USB 3.0
  4. Mode switch模式拨杆
  5. 10G SFP+
Front正面FIG. 06
USB‑C
Plug into any laptop.插上任何笔记本。Device mode. No drivers.从机模式,免驱动。
Switch
One flip.一拨。USB‑C or 10G. The screen tells you which.USB‑C 或 10G,屏幕告诉你。
10G SFP+
Serve the whole office.服务整个办公室。One model server for every device.一台服务器,所有设备。
Agent Ready

Models preloaded.
Ready at power‑on.
预装基础模型,
开机即用。

MaixStation top view with its display
  • vLLM
  • SGLang
  • PyTorch
  • llama.cpp
  • Ollama
  • OpenClaw
  • PicoClaw
  • Hermes Agent
  • Claude Code
  • Codex
  • n8n
  • MCP
Open a browser and go.打开浏览器就能用。Qwen3.8‑Flash‑Next and a web UI, out of the box. DGX Spark ships empty.Qwen3.8‑Flash‑Next 和 Web UI 出厂自带。DGX Spark 开箱是空的。
APIs on the device.API 跑在设备上。OpenAI and Anthropic compatible. Any agent, any phone browser.OpenAI 和 Anthropic 兼容。任何 agent,任何手机浏览器。
Your stack, unchanged.你的技术栈,照常用。vLLM, SGLang, PyTorch, llama.cpp and Ollama, NPU‑accelerated.vLLM、SGLang、PyTorch、llama.cpp 和 Ollama,NPU 加速。
Affordable

32GB, 128GB, 512GB.
No markup. Pick what fits.
从 32G 到 128G 到 512G,
不再溢价,丰俭由人。

$8k$6k$4k$2k$0 32GB64GB128GB256GB UNIFIED MEMORY · LOG SCALE $4,999DGX Spark 64GB $6,999DGX Spark 128GB $1,xxxMaixStation 32GB from $3,xxxMaixStation 128GB

Past two units, DGX Spark needs a 200G/400G switch. About $4,000 more. MaixStation needs a cable.超过两台,DGX Spark 需要 200G/400G 交换机,多花约 $4,000。MaixStation 只需要一根线。

Scalable

One, two, four,
or more. Your call.
一台,两台,四台,
及更多,都随你。

Every tier is a cable away. No expensive switch.每一级只隔一根线。无需昂贵交换机。

1 × NODE 32GB Qwen3.8‑27B 1 × NODE 128GB Qwen3.8‑Flash‑Next 125B 2 × NODE MCIO 8i · PCIe 5.0 ×8 256GB DeepSeek V4‑Flash 284B MAIXSTATION PRO 512GB DeepSeek V4.1‑Flash 552B AND BEYOND 512GB 1 × MaixBlade · PCIe 5.0 ×16 DeepSeek V4.1‑Flash 552B 1TB 2 × MaixBlade · PCIe 5.0 ×16 DeepSeek V4‑Pro 1.6T 2TB 4 × MaixBlade · PCIe 5.0 ×16 Kimi K3 2.8T
MaixBlade compute card
MaixBlade · 4 × AX8860P, PCIe card四芯 PCIe 算力卡FIG. 03
Maix-V core module with the AXERA AX8860P
Maix‑V · AX8860P · 78 × 78 mm module78 × 78 mm 核心板FIG. 04
Scenarios场景

What it does.它能做什么。

Video generation视频生成

Generate video locally. No per‑second fees.本地生成视频,不按秒计费。

MaixStation 128GB · open video models

Text‑to‑video and image‑to‑video run on your own box. No queue, no cloud bill, footage never leaves your desk.文生视频、图生视频跑在自己的机器上。不排队,没有云账单,素材不出你的桌面。

Vibe codingVibe coding

Coding agents, unlimited tokens.写代码的 agent,token 不限量。

MaixStation 128GB · Qwen3.8‑Flash‑Next 125B

Point Claude Code, Codex or OpenClaw at the on‑device API. Flat cost.把 Claude Code、Codex、OpenClaw 指向设备上的 API,成本固定。

On a robot机器人上

VLA runs on the robot itself.VLA 直接跑在机器人上。

MaixStation 32GB · VLA

Real‑time policies at a fraction of Jetson‑class cost.策略实时执行,成本远低于 Jetson 级方案。

In your bag包里

A 27B model, on the go.出门也能跑 27B。

MaixStation 32GB · Qwen3.8‑27B

Full 27B on a power bank. Plug into any laptop.满血 27B,充电宝供电,插上笔记本就能用。

On your desk办公桌上

Office work, handled automatically.办公室,自动处理。

MaixStation 128GB · Qwen3.8‑Flash‑Next 125B

Agents for documents, mail and data. Add a second unit over MCIO when you want DeepSeek V4‑Flash 284B.处理文档、邮件和数据的 agent。想跑 DeepSeek V4‑Flash 284B 时,MCIO 加一台。

In the company公司里

One model for everyone.一个模型,全员可用。

MaixStation Pro / MaixBlade · 512GB+

DeepSeek V4.1‑Flash 552B for the whole office. Stack boxes or cards as you grow.DeepSeek V4.1‑Flash 552B 服务全公司。一体机或算力卡按需级联。

FAQ
When does it launch?什么时候上线?
Kickstarter, end of 2026. Reservation holders get the date first and 24 hours of priority on the Super Early Bird tier.Kickstarter,2026 年底。预缴意向金的用户第一时间收到日期,并享有 Super Early Bird 档位 24 小时优先。
What chip is inside?什么芯片?
AXERA AX8860P. 256 TOPS FP4, 256‑bit LPDDR5X‑8533, 273 GB/s. On a 78×78 mm Maix‑V module with an LPCAMM2 connector, shared by MaixStation, MaixStation Pro and MaixBlade.爱芯元智 AX8860P。256 TOPS FP4,256‑bit LPDDR5X‑8533,273 GB/s。位于 78×78 mm 的 Maix‑V 核心板,LPCAMM2 接口,MaixStation、MaixStation Pro 和 MaixBlade 共用。
Can I run a model you haven't converted?没转过的模型能跑吗?
Yes. The conversion toolchain is open. Hugging Face checkpoint in, deployable model out. We ship converted builds of the major open models on release day; the tool is yours for the rest.可以。转换工具链开放。Hugging Face 权重进,可部署模型出。主流开源模型发布当天提供转换版本,其余用同一套工具自己转。
Why no 64GB?为什么没有 64GB 版本?
No flagship open model lands there. 32GB runs the best dense 27B; 128GB runs the best 125B MoE. 64GB would run neither better. Enterprise customers who need it can order a custom build.没有与之匹配的高性能模型。32GB 跑最好的稠密 27B,128GB 跑最好的 125B MoE,64GB 两边都不占。有企业需求可以定制。
Buying in volume?需要批量采购,怎么联系商务?
Email support@sipeed.com with your quantity and use case, and our business team will get back to you.请发送邮件到 support@sipeed.com,注明采购数量和使用场景,商务团队会尽快与你联系。
Reserve锁定优惠

Pay $30 now.
$100 off when you buy.
预缴 $30。
购买抵 $100。

  1. 01Answer a short survey to lock your spot.填一份简短问卷,锁定名额。
  2. 02Pay the $30 reservation fee when it opens and get a voucher code (one per machine).意向金开放后预缴 $30,拿到 F 码(一机一码)。
  3. 03Back us on Kickstarter and message us the code: $100 off, refunded to your pledge.在 Kickstarter 下单后私信我们 F 码,抵 $100(退回到你的 Kickstarter 订单)。
Lock your spot去锁定名额 →Reservation fee opens soon.意向金即将开放。
Community社区

Follow the build.关注进展。