Token 自由在加速：Redis 之父让你的 Mac 跑起 DeepSeek V4 Flash
Token Freedom Accelerating: Redis Creator Runs DeepSeek V4 Flash Natively on Mac
结论先行（BLUF）：Redis 创始人 antirez 写了一个 DeepSeek V4 Flash 的 Mac 专用推理引擎 ds4，40,000 行 C 代码，17 个 Metal Shader，只为一个模型——284B 参数的 MoE 巨兽。128GB 内存跑 2-bit 量化版，make 一下就能启动，OpenAI / Anthropic 双 API 兼容。这不是玩具。
为什么值得关注
当大多数人在用 Ollama 包装通用运行时的时候，antirez 写了一个只跑这一个模型的引擎。
理由很简单：DeepSeek V4 Flash 是 284B 参数的 MoE 架构，激活参数 13B，上下文窗口 1M token。要在 Mac 上把它跑好，通用方案的每一处妥协都是代价。ds4 的回答是：把通用性全部砍掉，把对这一个模型的理解吃透，然后榨干 Apple Silicon 的每一分性能。
这就是写出 Redis 的那个人的做事方式。
三个核心工程决策
1. 非对称 2-bit 量化——聪明地偷懒
不是所有参数都平等对待。ds4 只压缩路由专家（Routed MoE Experts），用 IQ2_XXS / Q2_K 极限量化；**共享专家（Shared Experts）**保持原始精度。
结果：整体模型压到 128GB 可装，但影响输出质量最大的那部分没有被委屈。这是对模型内部结构的理解，不是工具上默认的”全模型 2-bit”。
2. 磁盘 KV Cache——一等磁盘居民
“KV Cache 不该是内存的二等公民，它应该是一等磁盘居民。“——antirez
传统本地推理的痛点：服务重启 = 长上下文全部重新预填充。ds4 把 KV Cache 序列化到 SSD，包含 token ID、渲染文本和完整会话状态。重启服务，从断点恢复，无需重跑。
1M token 上下文窗口在本地推理场景真正变得可用，靠的就是这个设计。
3. Metal 原生图执行器
17 个 Metal Shader 文件，为 Apple Silicon 的 GPU 架构量身编写。没有 PyTorch，没有 GGML 的通用层，直接操控硬件。CPU 推理路径仅用于正确性验证，ds4 的战场是 Metal。
实测性能
| 硬件 | 场景 | 预填充速度 | 生成速度 |
|---|---|---|---|
| M3 Max 128GB | 短提示 | 58.5 t/s | 26.7 t/s |
| M3 Max 128GB | 11,709 token 长提示 | 250 t/s | 21.5 t/s |
| M3 Ultra 512GB | — | 比 M3 Max 快 44–78% | — |
M3 Ultra 上接近 468 t/s 预填充（用户提供数据）。对于 284B 参数的模型，这是实用级别的速度，不是跑来截图的玩具数字。
上手三步
# 1. 下载模型（2-bit，约 128GB）
./download_model.sh q2
# 2. 编译
make
# 3a. 命令行交互
./ds4
# 3b. 启动带磁盘 KV Cache 的 HTTP 服务
./ds4-server --ctx 100000 --kv-disk-dir /tmp/ds4-kvHTTP 服务兼容 OpenAI 和 Anthropic API 格式，Claude Code 可以直接在 MCP 配置里接入，不需要任何云端 API Key。
常见问题
Q：128GB 够跑吗？
A：够，专门为 128GB 设计了 2-bit 非对称量化版本（q2）。256GB+ 可以跑 4-bit 版本，质量更高。
Q：和 Ollama / llama.cpp 有什么区别？
A：ds4 不是通用运行时，只跑 DeepSeek V4 Flash。好处是针对这个模型的每一处优化都可以做到极致，不需要考虑对其他模型的兼容性。
Q：磁盘 KV Cache 会影响生成速度吗？
A：不影响生成阶段速度，影响的是从断点恢复的时间（远比重新预填充快）。生成阶段仍然全部在内存中完成。
Q：能在 Intel Mac 上跑吗？
A：不能，Metal 路径专为 Apple Silicon 设计。CPU 推理路径存在，但 antirez 明确标注有 macOS 虚拟内存 bug，不建议用于实际推理。
Q：Claude Code 怎么接入？
A：启动 ./ds4-server，在 Claude Code 的 MCP 配置里添加 baseURL: http://localhost:PORT，模型名填对应格式即可。完全本地，零云端依赖。
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。
BLUF: Redis creator antirez just open-sourced ds4 — a Mac-native Metal inference engine written from scratch for one model: DeepSeek V4 Flash. 40,000 lines of C, 17 Metal shaders, asymmetric 2-bit quantization, disk-persistent KV cache. Runs on 128GB RAM. One make command. OpenAI and Anthropic API compatible.
- GitHub: github.com/antirez/ds4
- Announcement: x.com/antirez/status/2052405820235678175
Why This Matters
While most local inference tools wrap general-purpose runtimes, antirez built an engine for one model only.
DeepSeek V4 Flash is a 284B-parameter MoE model with 13B active parameters and a 1M-token context window. Running it well on a Mac means every compromise in a general-purpose runtime has a real cost. ds4’s answer: eliminate all generality, deeply understand this one model, then extract every bit of performance from Apple Silicon.
This is how the author of Redis approaches engineering.
Three Core Engineering Decisions
1. Asymmetric 2-bit Quantization — Smart Laziness
Not all parameters are treated equally. ds4 aggressively quantizes routed MoE experts (IQ2_XXS / Q2_K) while keeping shared experts at full precision.
Result: the full model fits in 128GB, but the components that most affect output quality aren’t compromised. This requires understanding the model’s internal structure — not just running “global 2-bit” from a tool’s defaults.
2. Disk KV Cache — First-Class Disk Citizen
“KV cache shouldn’t be a second-class memory citizen — it should be a first-class disk citizen.” — antirez
The classic local inference pain point: server restart = re-prefill everything. ds4 serializes KV cache to SSD, including token IDs, rendered text, and full session state. Restart the server, resume from the checkpoint, no reprocessing needed.
A 1M-token context window becomes practically usable locally because of this design.
3. Metal-Native Graph Executor
17 Metal shader files, written specifically for Apple Silicon GPU architecture. No PyTorch, no GGML abstraction layer — direct hardware control. The CPU inference path exists for correctness validation only; Metal is the real target.
Performance Numbers
| Hardware | Scenario | Prefill Speed | Generation Speed |
|---|---|---|---|
| M3 Max 128GB | Short prompt | 58.5 t/s | 26.7 t/s |
| M3 Max 128GB | 11,709-token prompt | 250 t/s | 21.5 t/s |
| M3 Ultra 512GB | — | 44–78% faster than M3 Max | — |
For a 284B-parameter model, these are production-usable numbers — not just screenshot benchmarks.
Getting Started
# Download 2-bit model (~128GB)
./download_model.sh q2
# Build
make
# CLI mode
./ds4
# HTTP server with disk KV cache
./ds4-server --ctx 100000 --kv-disk-dir /tmp/ds4-kvThe HTTP server is compatible with both OpenAI and Anthropic API formats. Claude Code can connect directly via MCP config — no cloud API key required.
FAQ
Q: Is 128GB really enough?
A: Yes — the q2 build uses asymmetric 2-bit quantization designed specifically for 128GB systems. 256GB+ can run the 4-bit variant for higher quality.
Q: How is this different from Ollama or llama.cpp?
A: ds4 is not a general-purpose runtime — it only runs DeepSeek V4 Flash. That constraint allows every optimization to be pushed to the limit without worrying about other model compatibility.
Q: Does disk KV cache slow down generation?
A: No impact on generation speed. It only affects recovery time from checkpoints (which is far faster than re-prefilling). Generation still runs fully in memory.
Q: Does it work on Intel Macs?
A: No — the Metal path is Apple Silicon only. A CPU inference path exists, but antirez explicitly notes macOS virtual memory bugs that make it unsuitable for real inference.
Q: How do I connect Claude Code to ds4?
A: Start ./ds4-server, then add baseURL: http://localhost:PORT in Claude Code’s MCP config with the appropriate model name. Fully local, zero cloud dependency.
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
💬 评论与讨论
使用 GitHub 账号登录后发表评论