☰ Menu
Home›AI & Machine Learning›Strata
App Open source

Strata for Mac

Strata

Run powerful AI on your own computer. Chat, code and work with images, locally.

Strata — overview

Run powerful AI on your own computer. Chat, code and work with images, locally.

Models · Install · Hardware · Tools · Help · Support

English · 简体中文 · 日本語 · Deutsch · Français · Español · Português · Polski

Free, open source, and yours to run. Use Strata on your desktop or with your favorite coding tools. No account needed.

Models

We recommend ISTA-DASLab first, then Unsloth and other providers, where a compatible model fits.

 Qwen3.8-Flash-Next

Compare sizes, memory needs and compatibility.

 Qwen3.6-35B-A3B

Unsloth: UD-Q4_K_XL, UD-Q6_K_XL, Q8_0. These are the currently documented packs for this model.

Model guide and compatibility.

 DeepSeekV4.1-Flash-1017

In the desktop app and the Terminal. What does not fit in memory is read from your SSD, so it needs a lot of free disk space.

Runs on NVIDIA RTX 30, 40 and 50 series (Windows and Linux), AMD Radeon AI PRO R9700 (Linux), AMD Ryzen AI Max — Strix Halo (Windows), and Apple Silicon Macs with 64 GB or more (Mac). Setup downloads the files and the DeepSeek engine, then starts the model.

Measured with the 2.00 bpw set on Strata 0.2.5 (Strix Halo: 0.2.7). Output is the speed of the answer, prompt is how fast Strata reads your input.

Memory, disk and details · Release notes

Not sure what fits? Choose a model for your computer.

Install

Desktop app

Download Strata for Windows, Linux or Mac →

Install the app, choose a model, and start chatting. There is one installer for each system. On first start the app downloads the engine for your graphics card. Help with the installer.

Files: Strata-0.2.13-windows-x64-setup.exe · Strata-0.2.13-linux-x86_64.AppImage · Strata-0.2.13-macos-arm64.dmg

Chat, files and a live preview in one window.

Terminal app

Prefer the terminal? Install Strata, then run strata launch.

Install commands

Linux / Mac

Windows PowerShell

On Windows, open a new terminal afterwards. Then:

Follow the prompts to download and start a model. Terminal guide.

Watch your model and hardware while it runs.

Build from source · Docker · Full install guide

Hardware

Tools

Claude Code · Codex CLI · OpenCode · oh-my-pi · pi.dev · Hermes Agent · OpenClaw, and more through the OpenAI- and Anthropic-compatible APIs.

Run strata connect to see setup instructions for your tool. Connection guide.

Help

All guides · Troubleshooting · Report a bug · Benchmarks · How it works · Contribute

Support Strata

Help keep local AI open to everyone. A GitHub sponsorship or a coffee helps fund development and testing on more hardware.

Loyal supporters

Thank you to everyone who supports Strata.

MIT license · Credits and model licenses · Privacy

For AI tools and automated readers: technical reference · setup instructions.

Sponsorships, hardware partnerships or collaborations: niko.veit@outlook.com.

Details

以前在 8GB 显卡下，开启 64K 上下文会耗尽专家缓存空间，导致启动失败。只能改用 CPU 图像处理、更小的 KV 缓存，或将上下文降到约 48K。

Strata 0.2.10 改进了这套机制：当显存不足时，优先调整显存和内存之间的数据分配，尽可能保留用户设置的上下文长度。

比如，现在可以在 8GB 显存条件下启动 Qwen3.8 IQ3_S，并保持 128K 上下文，也就是 131,072 Token。

Strata 会根据显存余量，逐步缩小驻留在 GPU 中的 KV 缓存窗口，必要时还会调整其他缓存，甚至将部分数据放到系统内存中。

官方测试中：

  • 8GB 显存限制下，DFlash2 模式生成速度达到 63～73 Token/s。
  • 32K Token 的提示词处理时间为 19.8 秒。
  • 一张真正的 Intel Arc A750 8GB 显卡，也成功启动 128K 上下文，生成速度为 12.7～14.8 Token/s。

注意，显存需求仍然在，不是凭空消失，而是把压力给到了内存和 SSD。

Strata 0.2.10 将 DFlash2 设置为默认加速器。

DFlash2 是一种推测解码器。简单理解，就是让一个体积较小的辅助模型提前猜测接下来可能生成的内容，再交给主模型批量验证。猜对的部分越多，主模型就越有机会一次确认多个 Token，从而加快生成速度。

此前 Strata 使用的 MTP 推测层大约需要 5.2GB 空间，而新版 DFlash2 的显存占用约为 625MiB。

官方对比了同一引擎下 DFlash2 和 MTP 的表现：

测试设备 模型 生成速度提升
Ryzen AI Max（Windows） IQ3_S 短回复 +8.4%
RTX PRO 5000 IQ3_S 短回复 +6.23%
Apple M5 Max IQ3_S +3.8%
DGX Spark IQ3_S +5.8%

但是，也不是没有代价。

DFlash2 与 MTP 对比

相比之下，MTP 在硬件兼容性和多并发下都有优势：

对比项目 DFlash2 MTP
显存占用 约 625MiB 约 5.2GB
额外文件 首次下载约 1.2GB 推测层约 5.2GB
单请求生成速度 多数官方测试更快 部分硬件更快
多并发 主要面向单请求 更适合多请求
低显存设备 优势明显 显存压力较大
支持范围 依赖特定模型与引擎 兼容范围更广
Strata 0.2.10 默认策略 单请求 Qwen3.8 优先 多并发等场景优先

两张显卡一起跑，代码编辑速度接近翻倍

针对多显卡，这次更新带来优化：All-cards Split Engine。

以前多显卡运行模型时，常见做法是将模型的不同层分配到不同显卡。新的方式则允许多张显卡共同处理每一层模型，让它们同时参与计算。

官方使用两张 RTX A6000，在 Linux 下运行 UD-Q4_K_XL，得到以下结果：

测试任务 原来的速度 新版速度
代码编辑 195 Token/s 377～394 Token/s
生成新代码 139～144 Token/s 189～232 Token/s
普通文本 93 Token/s 139～146 Token/s
Agent 循环 112～119 Token/s 165 Token/s

代码编辑任务的生成速度几乎翻倍。

还加入了一个编程专用的 DeepSeek 模型

Strata 0.2.10 新增了 ISTA-DAS Coder。它基于 DeepSeek-V4.1-Flash，针对编程和 Agent 任务进行了专家裁剪，每层只保留原来 384 个专家中的 192 个。

这样可以减少需要处理的专家权重。

官方测试中：

  • 单张 Radeon AI PRO R9700：22.8 Token/s。
  • 四张 R9700：59.2 Token/s。
  • Apple M5 Max：19.1～22.5 Token/s。
  • RTX PRO 6000 Blackwell：约 100 Token/s。

How to install Strata

  1. Open Terminal (press ⌘ + Space and type “Terminal”).
  2. If Homebrew is not installed yet, run the command below first.
  3. Run this command:

Latest 10 releases

Last updated 2026-11-09; current version v0.2.13, with the latest 10 releases listed.

VersionReleasedTypeSizemacOSWindowsLinux
v0.2.132026-11-09Stable161.4 MBDownloadDownloadDownload
v0.2.122026-11-07Stable161.1 MBDownloadDownloadDownload
v0.2.112026-11-05Stable161.0 MBDownloadDownloadDownload
v0.2.102026-11-02Stable160.8 MBDownloadDownloadDownload
v0.2.92026-10-31Stable160.3 MBDownloadDownloadDownload
v0.2.82026-10-29Stable160.2 MBDownloadDownloadDownload
v0.2.72026-10-27Stable160.2 MBDownloadDownloadDownload
v0.2.62026-10-25Stable159.0 MBDownloadDownloadDownload
v0.2.52026-10-25Stable159.1 MBDownloadDownloadDownload
v0.2.42026-10-24Stable158.7 MBDownloadDownloadDownload

Data from GitHub Releases; stable releases first, each group sorted newest first.

Compared with similar apps

Only public metadata is compared, with no subjective recommendation; see the compare pages for field-by-field comparisons.

Compared with similar apps
AppVersionLicenseArchitecture30-day installs
Stratav0.2.13MITUniversal binary0
Codexrust-v0.169.0—Universal binary417,825
ollamav0.40.16MITCLI63,182
Claude Code2.1.308—Universal binary60,329

Similar apps: AI & Machine Learning

Codex

Codex CLI is a coding agent from OpenAI that runs locally on your computer.AppOpen source417,825/mo

ollama

Create, run, and share large language models (LLMs)CLIOpen source63,182/mo

Claude Code

Claude Code is an AI programming assistant from Anthropic that turns the terminal and IDE into…App60,329/mo

llama.cpp

ggml / ops / maintainer PRs%20sort%3Aupdated-desc) / dev stats / lib llama API / llama-server…CLIOpen source34,660/mo

ggml

Tensor library for machine learningCLIOpen source27,342/mo

rtk

CLI proxy to minimize LLM token consumptionCLIOpen source23,446/mo

Apps under the same license

Categories: AI & Machine Learning

Leave a comment

评论表单加载中…