Run DeepSeek V4 Flash 284B on 64GB RAM? llama-modchip Reality Check

Last updated: November 9, 2026

If you searched for how to run DeepSeek V4 Flash on 64GB of RAM, the short answer is: one new repo says you can, and the Hacker News thread says to check before you trust it. The project is llama-modchip, it hit the front page on November 8, and it makes a bold claim: a 284 billion parameter mixture-of-experts model running on a CPU-only box with 64GB of memory.

This post separates what the repo states from what the community could reproduce, and what a builder should do with it on Monday.

What the repo says#

The README describes llama-modchip as a set of ggml and llama.cpp kernel patches, quantization transcoders and runtime tools for running sparse MoE checkpoints in system RAM. These are the author's figures, not independent measurements:

• The model. DeepSeek-V4-Flash-0731 with 284.33B total parameters and 13.17B active per token. The standard UD-IQ1_S build is listed at 76.87 GiB, which does not fit in about 62.5 GiB of usable Linux RAM.
• The squeeze. A custom "Fractal" checkpoint at 56.56 GiB and 1.71 bits per weight, with 57.61 GiB peak resident memory at 256K context.
• The speed (vendor-reported). About 44 tokens per second prefill and 9.06 to 9.95 tokens per second pure decode on a 16-core AMD Ryzen 9 9955HX with no GPU. A self-speculative n-gram drafter lifts decode to 11.5 to 14.7 tokens per second in the tables.
• The smaller model. A 30B-A3B Nemotron build at 3.75 bits per weight, listed at 25 to 60 tokens per second on the same CPU.
• Agent hooks. KV cache slot snapshots in /dev/shm that restore in 13 to 33 ms for a 40-tool Claude Code style prompt.

The README says the default "lossless" mode evaluates all six routed experts and matches the unpatched runtime on logit cosine similarity. Faster modes drop experts (Top-5, Top-4) and the repo itself reports small perplexity and top-1 match losses for those.

The command, and where it breaks#

The README quickstart for the large model is two steps (copied from the repo; I did not run it):

hf download Synchopate/DeepSeek-V4-Flash-0731-Fractal-64GB-CPU-GGUF --local-dir ./dsv4
./dsv4/llama-server-modchip-linux-x86_64 -m ./dsv4/DeepSeek-V4-Flash-0731-Fractal-00001-of-00003.gguf -c 32768 -t 16 -fa on --port 8080

The compatibility note matters more than the command. The checkpoint uses a custom block-64 tensor type (type ID 40), so unpatched llama.cpp, Ollama and LM Studio will reject it with "unknown tensor type 40". If you already run a stack from our runtime comparison, this does not slot in. It is a separate binary.

What people are actually saying#

The Hacker News thread reached roughly 440 points, and the mood was sceptical:

• Reproduction failures. One commenter on a 24-core Xeon with 96GB reported the setup defaulting to Nemotron, then a tensor shape error when loading the DeepSeek file, and an "illegal instruction" crash from the bundled binary. Another with a Ryzen 7945HX and 96GB said Nemotron ran about the same as plain llama.cpp except for faster prefill, and called the repo AI-generated.
• Speed in practice. A commenter reported 116 seconds to process 43 input tokens and answer a simple translation request, and compared it with under two seconds on a GPU rig running a smaller model. They concluded it is one coding task per night.
• Quality doubts. A commenter who had used the full 0731 model said anyone claiming good quality at that size should be doubted, and that a 284B model needs far more memory to run without discarding information. Another wondered which baseline file the 76.87 GiB figure refers to, since the published Unsloth quants are larger.
• The counter-case. Others pushed back that big models degrade less under heavy quantization than people assume, and pointed to a different published quant that reports parity with FP8 at 2 bits. A third asked the fair question: bad compared to the original, but is it bad compared to what else fits in 64GB?
• Economics. A hosted DeepSeek Flash plan, one commenter argued, gives a lot of agentic use for about $20 a month, and you keep moving to newer models as they ship.

The pattern is familiar from our earlier look at open-weight models and the HN reaction: the ecosystem energy is real, and the specific headline claim is where the arguing happens.

The angle: who wins and loses#

Who this helps. People who must keep code on their own hardware and already own 64GB of fast DDR5. For them the interesting result is not the 284B headline. It is the repo's smaller Nemotron 30B-A3B numbers and the idea of restoring a KV snapshot instead of re-reading a 40-tool prompt every turn. Both are useful to anyone building a local agent loop.

Who it does not help. Anyone paying for hosted inference who values turnaround. Decode at roughly 10 tokens per second, plus a reasoning model that thinks at length, is slow for an agent that makes dozens of calls per task. Our DeepSeek V4.1 Flash release guide covers the hosted price and quality side, and the Ollama 0.40 MLX write-up shows the mainstream Mac path.

What to do Monday.

1. Run the repo's hardware readiness check, which needs no model download, before pulling 56 GiB.
2. Start with the 30B Nemotron model, not the 284B one. It fits 16 to 32GB and is the one commenters could run.
3. Measure end-to-end task time, not tokens per second. Reasoning length dominates on slow decode.
4. Compare against a hosted Flash model on the same prompts and keep whichever finishes the work you care about.
5. Do not use a 1.7 bits-per-weight model for anything where a silent quality drop would hurt, such as code you ship without review.

Continue Reading#

• DeepSeek V4.1 Flash 1017: Benchmarks, Pricing, Switch? - the hosted alternative
• Picking a Local Runtime for Coding Agents - Ollama, LM Studio, vLLM and llama.cpp compared
• Ollama 0.40 on Mac: MLX by Default - the supported Apple Silicon path
• Open-Weight AI's Kubernetes Moment - why the ecosystem argument keeps coming back
• GLM 5.3 free and cheap access - hosted open-weight options
• Best Local AI Models (2026) - open models ranked by hardware tier
• Where to Access AI Models (2026) - official APIs, routers and local weights side by side

Sources#

• llama-modchip on GitHub - README read in full November 9; all speeds and sizes above are vendor-reported
• Hacker News discussion - comments paraphrased from search excerpts of the thread
• daily.dev summary of the repo - independent restatement of the claims and community sentiment