Short answer: DeepSeek V4 Flash won’t run on a single consumer GPU, but it’s the most attainable 300B-class model to hit in months. The 304B-parameter MoE model needs about 82.5 GB of combined memory even at 1-bit quantization, roughly a third of what GLM-5.2 requires. Realistic options are a 24GB GPU plus 96GB of system RAM, a Mac Studio with 128GB+ unified memory, a dual-GPU workstation, or renting a single 80GB+ cloud card by the hour. This guide covers the exact numbers for each path.
Quick Navigation:
DeepSeek V4 Flash Architecture
DeepSeek V4 Flash 0731 landed at the end of July 2026 and immediately hit the top of Hacker News, driven by its ARC-AGI benchmark results: 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2 at max reasoning effort, both near frontier-model territory, at $0.02 and $0.04 per task on DeepSeek’s API. It ships MIT-licensed with full open weights.
DeepSeek V4 Flash 0731
DeepSeek hasn’t disclosed the exact activated-parameter count, only that it’s “far smaller” than DeepSeek-V4-Pro’s while beating it on benchmarks, which is the efficiency story behind the low API pricing. As with any MoE model, that efficiency applies to compute, not memory: all 304B parameters have to be loaded, since different tokens route to different experts. See our LLM quantization guide for why that gap between “active” and “total” params matters for hardware planning.
Memory Requirements by Quantization
These figures are for Unsloth’s dynamic GGUF quantizations (102 variants were live on Hugging Face within days of release). “Total memory” means VRAM plus system RAM combined, since llama.cpp can split MoE layers between GPU and CPU.
| Quantization | Total Memory Needed | Realistic Hardware |
|---|---|---|
| 1-bit (UD-IQ1_S) | 82.5 GB | 24GB GPU + 64-96GB RAM |
| 2-bit (UD-Q2_K_XL) | 96.8 GB | 128GB Mac Studio or GPU + 96GB RAM |
| 3-bit (UD-Q3_K_XL) | 128 GB | 2x GPU (96GB+ combined VRAM) |
| 4-bit (UD-Q4_K_XL) | 155 GB | 2x 80GB+ datacenter GPUs |
| 8-bit (UD-Q8_K_XL) | 162 GB | 2x H100/H200 |
Key insight: every quant level here is smaller than GLM-5.2’s minimum 223 GB. The 1-bit build is small enough that a single high-VRAM datacenter card, or a well-specced consumer GPU with generous system RAM, can actually hold it. That’s a meaningfully lower bar than most 300B+ open models set this year.
Four Ways to Run DeepSeek V4 Flash
Path 1: One 24GB GPU + 96GB System RAM
llama.cpp keeps attention layers and active experts on a single RTX 4090 or 5090 and streams the rest of the 1-bit quant’s MoE experts from system RAM. 96GB is enough headroom over the 82.5GB minimum for context and OS overhead. This is the cheapest entry point for a 300B-class model we’ve covered, see our RAM for local LLMs guide for the platform requirements.
Path 2: Mac Studio with 128GB+ Unified Memory
A Mac Studio with 128GB unified memory runs the 2-bit GGUF through llama.cpp’s Metal backend without any offloading complexity. It’s the simplest turnkey option and, unlike GLM-5.2’s 256GB requirement, doesn’t force you into Apple’s most expensive configuration.
Path 3: Dual-GPU Workstation (4-bit, Near-Lossless)
Two 80GB+ datacenter cards (2x A100 80GB or similar) give you 160GB+ of combined VRAM, enough to hold the 155GB 4-bit quant entirely in VRAM with no CPU offload penalty. This is the setup for anyone who wants near-full quality and fast, consistent throughput rather than the accuracy tradeoffs of 1-bit or 2-bit.
Path 4: Rent Cloud GPUs (Best for Trying It First)
A single rented 80GB+ card is enough to test the 1-bit or 2-bit quant before you buy anything, and a 2-4x 80GB node covers the near-lossless 4-bit build. Vast.ai usually has the lowest hourly prices on single and multi-GPU configurations, and RunPod gives new users a $5 credit to test with. See our cloud GPU comparison for current pricing.
DeepSeek V4 Flash vs GLM-5.2: Which Is Actually Self-Hostable?
Both are open-weight MoE models that made headlines this summer, but they sit on opposite ends of the self-hosting difficulty curve. GLM-5.2 is a 744B-parameter model that needs 223GB+ even at its smallest usable quant. DeepSeek V4 Flash is less than half the total size at 304B parameters, and its smallest usable quant comes in under 83GB, about a third of GLM-5.2’s floor.
For most home setups, that difference decides the question before you even look at benchmarks: GLM-5.2 realistically requires a 256GB Mac Studio, a large-RAM workstation, or serious multi-GPU spend. DeepSeek V4 Flash fits a much wider range of existing hardware, including rigs enthusiasts already own. Our GLM-5.2 hardware guide has the full breakdown if you’re weighing both.
Pro tip: whichever path you pick, quantize the KV cache. Running llama.cpp with –cache-type-k q4_1 –cache-type-v q4_1 stretches your usable context window roughly 3x at the same memory budget, especially useful given DeepSeek V4 Flash’s 384K max output length.
FAQ
Can you run DeepSeek V4 Flash on a single GPU?
Not the full model, but you’re closer than with most 300B+ models. The smallest usable quant (1-bit, UD-IQ1_S) needs about 82.5 GB of combined memory. A single 24GB GPU can’t hold that alone, but paired with 64-96 GB of system RAM for MoE offloading, or a single 80GB+ datacenter card, it fits.
How much VRAM does DeepSeek V4 Flash need?
The full BF16 model is about 167 GB. Quantized GGUF versions from Unsloth range from roughly 82.5 GB at 1-bit up to 162 GB at 8-bit. There’s no version that fits in 24 GB of VRAM alone; every practical setup either offloads to system RAM or spreads across multiple GPUs.
What’s the cheapest way to run DeepSeek V4 Flash locally?
One 24GB GPU (RTX 4090 or 5090) plus 96 GB of system RAM using llama.cpp MoE offloading, running the 1-bit quant. That’s a realistic enthusiast build, unlike 700B-class MoE models that demand 256 GB+ of RAM just to start.
Is DeepSeek V4 Flash easier to self-host than GLM-5.2?
Yes, significantly. GLM-5.2 needs around 245 GB even at 2-bit. DeepSeek V4 Flash’s smallest usable quant is under 83 GB, about a third of that. It’s the first 300B-class open-weight model in months that a well-specced single-GPU workstation can realistically run, even if only at 1-bit.
Why is DeepSeek V4 Flash getting so much attention?
It scored 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2, both near-frontier results, at $0.02 and $0.04 per task respectively, which is dramatically cheaper than comparable closed models. It’s also MIT-licensed with 102 community quantizations already live on Hugging Face within days of release.
Related Reading
Building for Big MoE Models?
Need a workstation that can hold 96GB+ of system RAM? Check our Tailored Builds page.