--- title: Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU | daily.dev description: This project claims to run the 284B-parameter DeepSeek-V4-Flash model on a 64GB CPU-only machine with under 59GiB RAM usage. It reports 44.36 tok/s prefill and... canonical: https://daily.dev/posts/llama-modchip-284b-deepseek-on-64gb-ram-no-gpu-wd0eeu1zf twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU | daily.dev og:description: This project claims to run the 284B-parameter DeepSeek-V4-Flash model on a 64GB CPU-only machine with under 59GiB RAM usage. It reports 44.36 tok/s prefill and... og:url: https://daily.dev/posts/llama-modchip-284b-deepseek-on-64gb-ram-no-gpu-wd0eeu1zf og:image: https://api.daily.dev/og/posts/WD0EEU1zf.png og:image:alt: Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU **[Hacker News](https://daily.dev/sources/hn)** · 16 min read · 2 upvotes · 0 comments ## Summary This project claims to run the 284B-parameter DeepSeek-V4-Flash model on a 64GB CPU-only machine with under 59GiB RAM usage. It reports 44.36 tok/s prefill and 9.1-14.7 tok/s decode at 256K context, using techniques it calls SwiGLU-H2 Hessian compression and Fractal-64 quantization. A 30B Nemotron MoE variant reportedly reaches 25-111 tok/s depending on hardware. It also claims a Claude Code integration with 13-32ms KV cache restore times. The models require a patched llama.cpp fork or a bundled pre-compiled binary since they use a custom tensor type. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How much RAM does it take to run a 284B parameter DeepSeek model on CPU only? Under 59 GiB of RAM at full 256K context, according to claimed benchmarks for DeepSeek-V4-Flash-0731-Fractal (284.33B total, 13.17B active parameters). The checkpoint is 56.56 GiB on disk, compressed from 76.87 GiB using techniques called SwiGLU-H2 Hessian compression and Fractal-64 (Q2_0, 3.75 bpw) quantization, reportedly reaching 44.36 tok/s prefill and 9.06-10.74 tok/s decode on a 16-core AMD Ryzen 9 9955HX with zero GPUs. _Anyone chasing GPU-free inference for huge MoE models can track projects like this on daily.dev._ ### Why do custom quantized GGUF models fail with 'unknown tensor type 40' in llama.cpp or Ollama? Tensor type 40 (GGML_TYPE_Q2_0) is being repurposed by custom block-64 BF16-scale Gray-coded tensors and custom deepseek4/Stockfish-PVS kernels that unpatched llama.cpp, Ollama, and LM Studio do not recognize. Running these 'Fractal-64' models requires either a patched llama.cpp build or the bundled pre-compiled standalone binary distributed alongside the model files on Hugging Face. _Developers debugging GGUF compatibility errors can follow quantization format changes like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 1 discussion and 91 comments across hackernews (as of 2026-11-08). **TL;DR:** Most skepticism centers on the claimed quantization level (believed to be ~2-bit), with many doubting output quality holds up despite the impressive token-rate numbers, while the README itself is widely criticized as unreadable AI-generated slop. **Sentiment:** 10% positive · 25% mixed · 65% skeptical **The case for** - A larger model at low quant can sometimes still beat a smaller model at higher quant, so aggressive quantization isn't automatically useless. - Some report that certain large models (e.g. DeepSeek variants) degrade surprisingly little even at 2-bit quantization. - The underlying achievement of fitting such a large model into 64GB is seen as an interesting technical direction worth exploring. **The pushback** - The quantization level isn't clearly disclosed and is suspected to be 2-bit or lower, which commenters associate with poor output quality. - The README is heavily criticized as AI-generated 'slop' that is confusing and nearly unreadable, making it hard to even verify the claims. - Several people note inconsistencies between claimed model sizes and known GGUF quant sizes, suggesting the technical writeup may be inaccurate or misleading. - Skeptics argue that running a heavily lobotomized model just to claim a feat isn't actually useful for real work. - General consensus that benchmarking tok/s instead of actual task quality is a red flag for sub-4-bit quant projects. **By community** - hackernews (skeptical): Heavy focus on doubting the actual quantization bit-depth and real-world usefulness, compounded by strong criticism of the AI-slop README, though a side thread on quantization theory and local-model setups is more neutral/technical. **Hottest debate:** Whether aggressive sub-4-bit quantization on large MoE models like this one actually preserves usable quality or just produces an impressive-sounding but lobotomized model. **Open questions** - What exact quantization scheme and bit-depth is actually used, since the readme's numbers don't match known GGUF file sizes. - How does output quality compare against existing well-known quantizations (e.g. unsloth's IQ2_XXS or IQ1_S) of the same base model. - What does 'MoE pruning' specifically entail in this implementation. **Highlights** > I’m confused too and the slop readme doesn’t help at all. > Shrinks DeepSeek-V4-Flash from 76.87 GiB (which crashes in OOM on 64 GB Linux) down to 56.56 GiB And later > Standard IQ2_XXS (76.87 GiB, Unpatched) Except unsloth’s IQ2_XXS GGUF is 91GB, even their smallest IQ1_S is 82.5GB. So I wonder which model/quant the number even refers to. Also the mentions of “compression” and “MoE pruning” also make it sound lobotomized even more than unsloth’s IQ1 quants. > — [adrian17 on hackernews · 1 comments](https://news.ycombinator.com/item?id=50412908) > PS: only now noticed that 76.87GiB _is_ 82.5GB, which happens to be size of the unsloth IQ1_S. The “technical” paper also mentions IQ1_S at some point. Sounds to me like the author’s model or the author mixed IQ2_XXS with IQ1_S at some point. Also, just noticed that “moe pruning” likely means “the model is supposed to run 6(DS)/8(Qwen) experts per token per layer, but we only run 4 to save on time”. Still doesn’t explain how they compressed the actual model, but I have no interest digging that info from that slop paper. > — [adrian17 on hackernews](https://news.ycombinator.com/item?id=50415077) > I very recently spent a lot of time using deepseek v4 flash 0731 before it was superseded by newer more capable things (glm 5.3-flash, v4.1 flash 1017, etc) and I am exceedingly skeptical of anyone that claims to get good QUALITY results out of it at that size. Whatever they're doing to shrink it will make it worse than an unsloth Q2XS quantization of it or equivalent. That size of model needs something like 150GB of RAM to run properly with full unquantized kv cache and not with a ton of training data discarded. I remain exceedingly skeptical of people who are claiming to run fairly large models in ridiculously tiny sizes. For comparison a 'regular' version of v4 flash 0731 has file sizes more like this: https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF Even the 97GB Q2XS will show very significant divergence from the full precision thing. Running a lobotomized model just to say you're doing it is really not accomplishing anything special. > — [walrus01 on hackernews](https://news.ycombinator.com/item?id=50412889) > It looks like 2 bit. I'm really struggling to understand why this got so many upvotes. Colibri and others already covered this ground months ago. > — [jazzypants on hackernews](https://news.ycombinator.com/item?id=50414239) > Anyone else finding these AI generated pages really hard to read? > — [spwa4 on hackernews · 1 comments](https://news.ycombinator.com/item?id=50413409) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50411961) · 244 points · 91 comments ## Similar posts on daily.dev - [MiniMax 2.5 vs Llama 3.1 vs DeepSeek: Local Coding Model Benchmark 2026](https://daily.dev/posts/minimax-2-5-vs-llama-3-1-vs-deepseek-local-coding-model-benchmark-2026-5pxrq2zot) · SitePoint · 1 upvotes · 0 comments - [DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Since Day 0](https://daily.dev/posts/deepseek-v4-1-flash-on-vllm-5x-agentic-throughput-since-day-0-js5lna7tm) · vLLM · 0 upvotes · 0 comments - [Serving DeepSeek-V4 on GB300 with SGLang: 5x Higher Throughput at the Same Interactivity Since Day-0 – PyTorch](https://daily.dev/posts/serving-deepseek-v4-on-gb300-with-sglang-5x-higher-throughput-at-the-same-interactivity-since-day-0-mmxalpv2n) · PyTorch · 1 upvotes · 0 comments --- Tags: [#python](https://daily.dev/tags/python), [#ai-agents](https://daily.dev/tags/ai-agents), [#claude-code](https://daily.dev/tags/claude-code) [View this post on daily.dev](https://daily.dev/posts/llama-modchip-284b-deepseek-on-64gb-ram-no-gpu-wd0eeu1zf) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU","url":"https://daily.dev/posts/llama-modchip-284b-deepseek-on-64gb-ram-no-gpu-wd0eeu1zf","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/llama-modchip-284b-deepseek-on-64gb-ram-no-gpu-wd0eeu1zf"},"datePublished":"2026-11-08T14:25:39.746Z","dateModified":"2026-11-08T18:56:59.838Z","description":"This project claims to run the 284B-parameter DeepSeek-V4-Flash model on a 64GB CPU-only machine with under 59GiB RAM usage. It reports 44.36 tok/s prefill and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ffc9d8e71444451f1da8fc300f78d61c?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/ffc9d8e71444451f1da8fc300f78d61c?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/llama-modchip-284b-deepseek-on-64gb-ram-no-gpu-wd0eeu1zf","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"python,ai-agents,claude-code","timeRequired":"PT16M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/llama-modchip-284b-deepseek-on-64gb-ram-no-gpu-wd0eeu1zf#faq","mainEntity":[{"@type":"Question","name":"How much RAM does it take to run a 284B parameter DeepSeek model on CPU only?","acceptedAnswer":{"@type":"Answer","text":"Under 59 GiB of RAM at full 256K context, according to claimed benchmarks for DeepSeek-V4-Flash-0731-Fractal (284.33B total, 13.17B active parameters). The checkpoint is 56.56 GiB on disk, compressed from 76.87 GiB using techniques called SwiGLU-H2 Hessian compression and Fractal-64 (Q2_0, 3.75 bpw) quantization, reportedly reaching 44.36 tok/s prefill and 9.06-10.74 tok/s decode on a 16-core AMD Ryzen 9 9955HX with zero GPUs. Anyone chasing GPU-free inference for huge MoE models can track projects like this on daily.dev."}},{"@type":"Question","name":"Why do custom quantized GGUF models fail with 'unknown tensor type 40' in llama.cpp or Ollama?","acceptedAnswer":{"@type":"Answer","text":"Tensor type 40 (GGML_TYPE_Q2_0) is being repurposed by custom block-64 BF16-scale Gray-coded tensors and custom deepseek4/Stockfish-PVS kernels that unpatched llama.cpp, Ollama, and LM Studio do not recognize. Running these 'Fractal-64' models requires either a patched llama.cpp build or the bundled pre-compiled standalone binary distributed alongside the model files on Hugging Face. Developers debugging GGUF compatibility errors can follow quantization format changes like this on daily.dev."}}]} ```