Can I Run DeepSeek V4 Flash on 64GB RAM? Only at 1.7 Bits
Until this week the answer was no: the smallest standard GGUF of DeepSeek V4 Flash is 76.87 GiB, and 64GB of RAM cannot hold it. On 8 November a project called llama-modchip reached the top of Hacker News with a 56.56 GiB build that runs on a 64GB PC with no GPU. We read its README, its Hugging Face files and its issue tracker. The speed claims come from one machine, and the quality has not been measured against the real model.
Bottom Line
- Yes, but only with one custom build. llama-modchip packs the 284B model into a 56.56 GiB GGUF at 1.71 bits per weight. Its author reports a 57.61-58.9 GiB peak on a 64GB Linux PC, with no GPU.
- Every standard quant is too big. Unsloth’s smallest file for V4 Flash 0731, UD-IQ1_S, is 76.87 GiB. The README puts usable RAM on a 64GB Linux system at about 62.5 GiB.
- Speed (author’s numbers, one laptop): 43.16-44.36 tok/s prompt processing and 9.06-9.95 tok/s generation on a 16-core Ryzen 9 9955HX with 64GB of DDR5. The n-gram drafter raises generation to 11.5-14.7 tok/s.
- Nobody else has published a 284B run yet. The eight community benchmark issues filed on 8 November all used the 30B Nemotron model. The DeepSeek repo had 234 downloads that night.
- Quality is unmeasured. “Lossless” means bit-exact against the same 1.71-bit file. The project publishes no comparison with DeepSeek’s weights.
- Stock tools refuse the file. llama.cpp, Ollama and LM Studio stop with
unknown tensor type 40.
| Your 64GB machine | Best published route to DeepSeek’s 284B | Reported speed | Catch |
|---|---|---|---|
| CPU only, AVX-512 (Zen 4 / Zen 5) | llama-modchip, V4 Flash 0731, 1.71 bpw | 9.06-9.95 tok/s | Author’s machine only; quality unmeasured |
| CPU only, AVX2 only | llama-modchip, build from source | Not published | Prebuilt binary failed on a Haswell Xeon |
| 64GB + RTX 3060 12GB + ~1.1 TB NVMe | Promethevs, V4.1 Flash 1017, full weights | 4.4-7.4 tok/s | Author’s numbers; Windows build PC |
| 46-64GB + 12GB NVIDIA + NVMe | GSQ-RCO 2-bit in Strata, V4.1 Flash | 9-11 tok/s (RTX 3060) | Engine author’s numbers |
If you want a model that just works on 64GB, start with the best local LLM for 64GB RAM instead.
What Changed This Week
| Date | Event | Source |
|---|---|---|
| 2 Nov 2026 | llama-modchip v2.1.0 and v2.2.0 | GitHub releases |
| 3 Nov 2026 | v2.3.0: “DeepSeek-V4-Flash 284B in < 60 GiB RAM at 256K Context” | same |
| 4 Nov 2026 | v2.4.0: DeepSeek kernel work, “+68% Prefill, 9.1-10.1 tok/s Decode” | same |
| 5 Nov 2026 | Synchopate/DeepSeek-V4-Flash-0731-Fractal-64GB-CPU-GGUF created | Hugging Face |
| 6 Nov 2026 | Promethevs posted: “Full DeepSeek-v4.1-Flash-1017, locally, on 12GB of VRAM” | Hacker News |
| 8 Nov 2026 | ”Llama-modchip: 284B DeepSeek on 64GB RAM, no GPU” at 388 points, 174 comments | Hacker News |
The repository was created on 29 October 2026. It had 544 stars when we read it on 8 November.
The Arithmetic: Why 64GB Was “No” Until Now
DeepSeek’s model card lists V4 Flash at 284B total parameters and 13B activated. The 0731 config.json has 43 layers, 256 routed experts, 6 experts per token and 1 shared expert. The license is MIT.
At 13B active, the model reads only a small part of its weights per token. The problem is holding all 284B in memory. These are Unsloth’s file sizes for V4 Flash 0731, summed from the Hugging Face API on 8 November 2026:
| File | Size (GiB) | Size (GB) | Fits 64GB RAM? |
|---|---|---|---|
| Unsloth UD-IQ1_S | 76.87 | 82.54 | No |
| Unsloth UD-IQ1_M | 80.93 | 86.90 | No |
| Unsloth UD-IQ2_XXS | 84.62 | 90.86 | No |
| Unsloth UD-Q2_K_XL | 90.18 | 96.83 | No |
| llama-modchip Fractal (1.71 bpw) | 56.56 | 60.73 | Yes, per the author |
The modchip file on Hugging Face sums to 60,772,069,110 bytes, including two small Linux binaries. That matches the README’s 60.73 GB for the model.
How modchip makes it fit
The README describes three kinds of change, applied to Unsloth’s UD-IQ1_S file:
- Kept at higher precision: embeddings, the dense first layer, the 43 shared experts and part of the attention, in Q4_K, Q5_K or F32.
- Re-packed attention: some attention matrices go to a custom “Fractal-64” type (
GGML_TYPE_Q2_0, 3.75 bits per weight with 64-weight blocks). - The 256 routed experts: mostly IQ1_S at 1.5625 bits per weight. The routed experts are most of the model, so this sets the 1.71 average.
The custom tensor type is why stock tools fail. The README says unpatched llama.cpp, Ollama and LM Studio “will reject them with unknown tensor type 40.”
Speed: Every Number Published So Far
All of these numbers come from the llama-modchip README. We did not run the model.
| Machine | Setting | Prompt processing | Generation |
|---|---|---|---|
| Ryzen 9 9955HX, 64GB DDR5, no GPU | Stock UD-IQ1_S (76.87 GiB) | Out of memory or swap | 5.91 tok/s |
| same | modchip v2.4.0, all 6 experts (“Lossless”) | 43.16-44.36 tok/s | 9.06-9.95 tok/s |
| same | Top-6 + n-gram drafter | 43.16-44.36 tok/s | 11.5-14.7 tok/s |
| same | Top-4 “Turbo 10+“ | 43.16-44.36 tok/s | 10.12-10.74 tok/s |
Three details matter:
- One machine. The Ryzen 9 9955HX is a 16-core Zen 5 chip. Every 284B number traces back to the author’s run on it.
- The community runs are a different model. The README’s leaderboard has one DeepSeek row, the author’s. The issue tracker had eight new benchmark issues on 8 November. We opened all eight: they report the 30B Nemotron model (
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Q4_0), not DeepSeek. - The drafter helps repetitive text most. It is an n-gram drafter. It predicts tokens that already appeared, so code edits gain more than new prose.
The direct-run example in the README starts the server with -c 32768, a 32K context. The 256K memory figure is the author’s ceiling, not the default.
What “Lossless” Means Here
The README labels the default mode “Lossless.” Its table reports a logit cosine similarity of 1.00000 and a KL-divergence of 0.0000.
Read the table header. Those numbers are “measured relative to the unpatched Top-6 execution on the 56.56 GiB Fractal checkpoint.” So they show that the speed patches do not change the 1.71-bit file’s output. They say nothing about how far the 1.71-bit file is from DeepSeek’s own weights.
We searched the README and the DeepSeek deep-dive document for a perplexity, HumanEval or MMLU result against the original model. We found none. The “100/100 quality gate” in the README is for the 30B Nemotron files.
Hacker News readers raised the same point. One commenter who used V4 Flash 0731 heavily wrote that they are “exceedingly skeptical of anyone that claims to get good QUALITY results out of it at that size.” The same commenter noted that 0731 was “superseded” by newer releases such as V4.1 Flash 1017.
Will It Run on Your CPU?
The README lists kernels for AVX-512 and AVX2 on x86_64, plus ARM64. It names AMD Zen 5 parts (Ryzen 9 9955HX, Ryzen 9 9950X, EPYC 9005) with “single-cycle 512-bit AVX-512.”
One Hacker News user with a 24-core Haswell-era Xeon and 96GB of DDR4 tried the DeepSeek build on 8-9 November 2026. They reported two failures:
modchip startstopped withcheck_tensor_dims: tensor 'blk.0.ffn_gate_tid2eid.weight' has wrong shape.- The bundled
llama-serverbinary stopped withIllegal instruction (core dumped). They later concluded that “none of the precompiled binaries” work on a CPU with no AVX-512.
Another user, on a Ryzen 9 7945HX with 96GB of DDR5, found no DeepSeek option in modchip setup. The first user also wrote that the default app “just installs Nemotron.”
So the practical rule is simple. Use a Linux PC with a Zen 4 or Zen 5 CPU if you want the prebuilt binary. On an AVX2-only CPU, plan to build from the patches yourself.
How to Run It
These are the README’s own commands for the 284B model. Read the code before you run a prebuilt binary from a new project.
# Check your CPU first (no model download)
curl -sL https://raw.githubusercontent.com/synchopate/llama-modchip/main/scripts/check_cpu_readiness.py | python3
# Download the 56.56 GiB model and the prebuilt Linux x86_64 server
hf download Synchopate/DeepSeek-V4-Flash-0731-Fractal-64GB-CPU-GGUF --local-dir ./dsv4
chmod +x ./dsv4/llama-server-modchip-linux-x86_64
# Start the server: 32K context, 16 threads, flash attention
./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
Close other programs first. The model alone takes about 57 of the roughly 62.5 GiB that a 64GB Linux system can use.
The Other Routes to DeepSeek’s 284B
Promethevs takes the opposite approach. It runs the newer DeepSeek-V4.1-Flash-1017 “from all 510 GB of its original weights” on a Ryzen 9 3950X with 64GB of DDR4, an RTX 3060 12GB and two NVMe drives. Its author reports these speeds on that one PC:
| Prompt type | Cold | Warm (best case) |
|---|---|---|
| Code | 5.8 tok/s | 7.4 tok/s |
| Reasoning | 4.7 tok/s | 4.4 tok/s |
| Agent | 5.2 tok/s | 4.8 tok/s |
| Plain exact decoding, same PC | 2.14 tok/s | — |
The catch is disk. Its stores need about 1,009 GiB, so plan for about 1.1 TB of free NVMe. It also needs an NVIDIA RTX 30 or 40 series card with 12GB. Unlike modchip, it reports agreement against the exact model, word by word.
GSQ-RCO 2-bit files are the third route, also for V4.1 Flash with a 12GB NVIDIA card. We covered them in DeepSeek V4.1 Flash GSQ-RCO quants.
What to Buy for This
Read this section as “what the published runs used,” not as a promise of the same speed.
- The RAM: 64GB is the floor for every route above. A Crucial 64GB DDR5-4800 kit covers it. As of 6 November 2026, 64GB DDR5 kits started at $929.99, with a median of $1,262.49 (our RAM price check). CPU-only speed follows memory bandwidth, so faster DDR5 helps; this kit is the value tier, not the fast one.
- The CPU: the author’s numbers come from a Zen 5 chip with AVX-512. The Ryzen 9 9950X is the desktop Zen 5 part the README names. Nobody has published a 284B run on it yet.
- The GPU, if you add one: a 12GB NVIDIA card opens the two V4.1 Flash routes. The RTX 3060 12GB above is the card both authors used.
For the wider picture, see the hardware we recommend by tier.
Honest Caveats
- We did not run any of this. Every speed, size and memory figure is the project author’s claim, with the source named.
- The 284B speed is one machine. No independent run of the DeepSeek build existed when we read the issue tracker on 8 November 2026.
- The model is the older V4 Flash 0731. DeepSeek published V4.1 Flash 1017 on Hugging Face on 17 October 2026. modchip’s DeepSeek file is built from 0731.
- 1.71 bits is extreme. The routed experts are mostly 1.5625-bit. Expect a quality cost that the project has not measured.
- The binary is new and prebuilt. Run it in a container or on a test machine first. The project is ten days old.
- 64GB leaves little room. The model and cache use about 57-59 GiB at the author’s settings. A browser and an IDE beside it can push the system into swap.
Sources
- llama-modchip README and releases: sizes, bit widths, tensor types, speeds, “unknown tensor type 40”, CPU support, commands (read 8 Nov 2026)
- llama-modchip DeepSeek deep-dive: KL-divergence measured against the same checkpoint
- llama-modchip issues: eight benchmark issues on 8 Nov 2026, all on Nemotron 30B
- Synchopate/DeepSeek-V4-Flash-0731-Fractal-64GB-CPU-GGUF: file sizes, created 5 Nov 2026, downloads (HF API, 8 Nov 2026)
- unsloth/DeepSeek-V4-Flash-0731-GGUF: standard quant sizes (HF API, 8 Nov 2026)
- deepseek-ai/DeepSeek-V4-Flash model card and V4-Flash-0731 config.json: 284B / 13B, 43 layers, 256 experts, top-6, MIT license
- Hacker News: Llama-modchip: 388 points, quality doubts, the Haswell failure report (8-9 Nov 2026)
- Promethevs README and Hacker News post: V4.1 Flash 1017 on 64GB DDR4 + RTX 3060 12GB, speeds, disk needs (6 Nov 2026)
See Also
- Best local LLM for 64GB RAM: what runs well on 64GB at a normal quant
- DeepSeek V4.1 Flash GSQ-RCO quants: the 2-bit V4.1 Flash route for a 12GB NVIDIA card
- Can you run DeepSeek V4.1 Flash locally?: the full-precision memory math
- DeepSeek V4-Flash vs gpt-oss-120b: the two big MoE models for unified memory
- Run an LLM without a GPU: why memory channels set CPU-only speed
- Used EPYC servers for CPU-only MoE inference: eight memory channels instead of two
- Should I buy RAM now or wait?: the November 2026 DDR5 prices