{"slug": "deepseekv4ssd-deepseek-v4-flash-0731-on-an-m-series-mac", "title": "DeepSeekV4SSD: DeepSeek-V4-Flash-0731 on an M-series Mac", "summary": "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.", "body_md": "Inspired by [Turbo Fieldfare](https://github.com/drumih/turbo-fieldfare),\nDeepSeekV4SSD streams routed experts from SSD to run all 284 billion parameters\nof `DeepSeek-V4-Flash-0731`\n\non an M-series Mac with about 30 GB of memory.\n\nThese 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.\n\n| Test | Prefill | Decode | Peak memory |\n|---|---|---|---|\n| Codex request with 14,000 input tokens | 180 Tok/s | 6.5 Tok/s | 30 GB |\n| 4,096-token prompt with one output token | 144.53 Tok/s | — | 15.56 GiB |\n| Short prompt, second run in one runtime | — | 6.41 Tok/s | 15.05 GiB |\n\nThe first two rows were measured with runtime versions `v1.0.3`\n\nand `v1.0.2`\n\n,\nrespectively. Performance changes with the prompt, SSD speed, and cache state.\nSee the [validation record](/yanun0323/deepseek_ssd/blob/master/docs/VALIDATION.md) for the full test details.\n\n**Download the app → Open the app → Download the 167 GB full model → Start the\nserver → Chat in the app or connect Codex**\n\n- Download the latest\n`DeepSeekV4SSD-macOS-arm64.zip`\n\nfrom[GitHub Releases](https://github.com/yanun0323/deepseek_ssd/releases/latest). - Extract the ZIP and open\n`DeepSeekV4SSD.app`\n\n. - Select\n**Download Model**. The default installation includes DSpark and uses about 167 GB. You can stop the download and resume it later. - Select\n**Start Server** after the model is ready. - Use the chat in the app, or connect Codex with the configuration below.\n\nThe local server starts at `http://127.0.0.1:11434`\n\nby default.\n\n| Item | Requirement |\n|---|---|\n| Mac | Apple Silicon M-series Mac |\n| macOS | macOS 15 or later |\n| Unified memory | 64 GiB or more |\n| Free storage | About 172 GB (160 GiB) |\n| Model storage | A fast internal, Thunderbolt, or USB4 SSD |\n| Internet | Required to download the model and app updates |\n\nImportant\n\nDeepSeekV4SSD is experimental. Model weights are not included with the app. Keep the default local server address unless another device must connect.\n\nStart the server in DeepSeekV4SSD. Then add this configuration to\n`~/.codex/config.toml`\n\n:\n\n```\nmodel = \"deepseek-v4-flash-0731\"\nmodel_provider = \"deepseek-v4-ssd\"\nmodel_reasoning_effort = \"high\"\n\n[model_providers.deepseek-v4-ssd]\nname = \"DeepSeekV4SSD\"\nbase_url = \"http://127.0.0.1:11434/v1\"\nwire_api = \"responses\"\nrequires_openai_auth = false\n```\n\nRestart Codex after you save the file. The local address does not need an API\nkey. The provider settings must be in the user-level config file. See the\n[official Codex configuration reference](https://developers.openai.com/codex/config-reference/)\nfor more options.\n\n- The main model has 284B total parameters and about 13B active parameters per token.\n- Common tensors stay in unified memory.\n- Routed experts use checkpoint-native FP4 weights and stream from SSD when needed.\n- The runtime uses an FP8 KV cache and a bounded expert cache to control memory use.\n- The installed model is verified against the pinned checkpoint revision.\n\n- The main model uses about 145 GiB.\n- DSpark adds about 10.12 GiB and is included in the default download.\n- Installing DSpark does not enable it. Enable\n**Use DSpark** in the runtime settings when you want to test speculative decoding. - You can remove DSpark without reinstalling the main model.\n\nThe server supports these endpoints:\n\n`GET /healthz`\n\n`GET /v1/models`\n\n`POST /v1/responses`\n\n`POST /v1/chat/completions`\n\n`POST /v1/completions`\n\nThe Responses API supports Codex tools and OpenAI function tools. The client\nmust run each tool and send the result back to the server. Read the\n[API guide](/yanun0323/deepseek_ssd/blob/master/docs/API.md) for fields, examples, and current limits.\n\nThe app shows prefill speed, decode speed, token counts, memory use, SSD read speed, cache hit rate, first-token wait time, and completion time.\n\nInference 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.\n\n- The runtime supports only the pinned\n`DeepSeek-V4-Flash-0731`\n\ncheckpoint. - The runtime processes one generation request at a time.\n- Images, audio, logprobs,\n`response_format`\n\n, and`stop`\n\nare not supported. - Very long input and output need more KV cache memory.\n- Performance depends on SSD speed, input length, and cache state.\n\nRead the [runtime research](/yanun0323/deepseek_ssd/blob/master/docs/RUNTIME_RESEARCH_2026-08-07.md) and\n[implementation plan](/yanun0323/deepseek_ssd/blob/master/docs/IMPLEMENTATION_PLAN.md) for the model contract,\nruntime design, and measured engineering decisions.\n\nDeepSeekV4SSD is not affiliated with DeepSeek. Review the model terms before you download and use the model.\n\nThe DeepSeekV4SSD source code is available under the [MIT License](/yanun0323/deepseek_ssd/blob/master/LICENSE).\nModel weights are not included and remain subject to their own terms.", "url": "https://wpnews.pro/news/deepseekv4ssd-deepseek-v4-flash-0731-on-an-m-series-mac", "canonical_source": "https://github.com/yanun0323/deepseek_ssd", "published_at": "2026-08-10 02:22:57+00:00", "updated_at": "2026-08-10 02:35:59.724151+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-infrastructure", "ai-tools"], "entities": ["DeepSeekV4SSD", "DeepSeek-V4-Flash-0731", "yanun0323", "Apple M5 Pro", "MacBook Pro", "GitHub", "Codex", "DSpark"], "alternates": {"html": "https://wpnews.pro/news/deepseekv4ssd-deepseek-v4-flash-0731-on-an-m-series-mac", "markdown": "https://wpnews.pro/news/deepseekv4ssd-deepseek-v4-flash-0731-on-an-m-series-mac.md", "text": "https://wpnews.pro/news/deepseekv4ssd-deepseek-v4-flash-0731-on-an-m-series-mac.txt", "jsonld": "https://wpnews.pro/news/deepseekv4ssd-deepseek-v4-flash-0731-on-an-m-series-mac.jsonld"}}