DeepSeekV4SSD: DeepSeek-V4-Flash-0731 on an M-series Mac

DeepSeekV4SSD, an experimental app from developer yanun0323, streams routed experts from SSD to run all 284 billion parameters of DeepSeek-V4-Flash-0731 on an M-series Mac with about 30 GB of memory, achieving 180 Tok/s prefill and 6.5 Tok/s decode on a Codex request with 14,000 input tokens, measured on a MacBook Pro with Apple M5 Pro, 18 CPU cores, 20 GPU cores, and 64 GiB unified memory. The app, available on GitHub Releases, requires macOS 15 or later, 64 GiB unified memory, and about 172 GB free storage, and it supports local server endpoints for Codex integration.

Inspired by Turbo Fieldfare https://github.com/drumih/turbo-fieldfare , DeepSeekV4SSD streams routed experts from SSD to run all 284 billion parameters of DeepSeek-V4-Flash-0731 on an M-series Mac with about 30 GB of memory. These results were measured on a MacBook Pro with an Apple M5 Pro, 18 CPU cores, 20 GPU cores, and 64 GiB of unified memory. DSpark was disabled. | Test | Prefill | Decode | Peak memory | |---|---|---|---| | Codex request with 14,000 input tokens | 180 Tok/s | 6.5 Tok/s | 30 GB | | 4,096-token prompt with one output token | 144.53 Tok/s | — | 15.56 GiB | | Short prompt, second run in one runtime | — | 6.41 Tok/s | 15.05 GiB | The first two rows were measured with runtime versions v1.0.3 and v1.0.2 , respectively. Performance changes with the prompt, SSD speed, and cache state. See the validation record /yanun0323/deepseek ssd/blob/master/docs/VALIDATION.md for the full test details. Download the app → Open the app → Download the 167 GB full model → Start the server → Chat in the app or connect Codex - Download the latest DeepSeekV4SSD-macOS-arm64.zip from GitHub Releases https://github.com/yanun0323/deepseek ssd/releases/latest . - Extract the ZIP and open DeepSeekV4SSD.app . - Select Download Model . The default installation includes DSpark and uses about 167 GB. You can stop the download and resume it later. - Select Start Server after the model is ready. - Use the chat in the app, or connect Codex with the configuration below. The local server starts at http://127.0.0.1:11434 by default. | Item | Requirement | |---|---| | Mac | Apple Silicon M-series Mac | | macOS | macOS 15 or later | | Unified memory | 64 GiB or more | | Free storage | About 172 GB 160 GiB | | Model storage | A fast internal, Thunderbolt, or USB4 SSD | | Internet | Required to download the model and app updates | Important DeepSeekV4SSD is experimental. Model weights are not included with the app. Keep the default local server address unless another device must connect. Start the server in DeepSeekV4SSD. Then add this configuration to ~/.codex/config.toml : model = "deepseek-v4-flash-0731" model provider = "deepseek-v4-ssd" model reasoning effort = "high" model providers.deepseek-v4-ssd name = "DeepSeekV4SSD" base url = "http://127.0.0.1:11434/v1" wire api = "responses" requires openai auth = false Restart Codex after you save the file. The local address does not need an API key. The provider settings must be in the user-level config file. See the official Codex configuration reference https://developers.openai.com/codex/config-reference/ for more options. - The main model has 284B total parameters and about 13B active parameters per token. - Common tensors stay in unified memory. - Routed experts use checkpoint-native FP4 weights and stream from SSD when needed. - The runtime uses an FP8 KV cache and a bounded expert cache to control memory use. - The installed model is verified against the pinned checkpoint revision. - The main model uses about 145 GiB. - DSpark adds about 10.12 GiB and is included in the default download. - Installing DSpark does not enable it. Enable Use DSpark in the runtime settings when you want to test speculative decoding. - You can remove DSpark without reinstalling the main model. The server supports these endpoints: GET /healthz GET /v1/models POST /v1/responses POST /v1/chat/completions POST /v1/completions The Responses API supports Codex tools and OpenAI function tools. The client must run each tool and send the result back to the server. Read the API guide /yanun0323/deepseek ssd/blob/master/docs/API.md for fields, examples, and current limits. The app shows prefill speed, decode speed, token counts, memory use, SSD read speed, cache hit rate, first-token wait time, and completion time. Inference runs on your Mac. Prompts and generated text stay in the local runtime unless the connected client sends them elsewhere. The app uses the network to download the model, check for updates, and accept configured API requests. - The runtime supports only the pinned DeepSeek-V4-Flash-0731 checkpoint. - The runtime processes one generation request at a time. - Images, audio, logprobs, response format , and stop are not supported. - Very long input and output need more KV cache memory. - Performance depends on SSD speed, input length, and cache state. Read the runtime research /yanun0323/deepseek ssd/blob/master/docs/RUNTIME RESEARCH 2026-08-07.md and implementation plan /yanun0323/deepseek ssd/blob/master/docs/IMPLEMENTATION PLAN.md for the model contract, runtime design, and measured engineering decisions. DeepSeekV4SSD is not affiliated with DeepSeek. Review the model terms before you download and use the model. The DeepSeekV4SSD source code is available under the MIT License /yanun0323/deepseek ssd/blob/master/LICENSE . Model weights are not included and remain subject to their own terms.