DeepSeek V4-Flash, a 284-billion-parameter AI model, can now run entirely on local machines without experimental software forks, thanks to mainline llama.cpp support that arrived in July 2026. The model requires at least 110GB of memory and a 103GB file size at recommended settings, putting it within reach of high-end consumer hardware and workstations for the first time.

What Changed Between June and July 2026?

In June, running DeepSeek V4-Flash locally was a technical obstacle course. Every consumer deployment relied on experimental forks of llama.cpp, the open-source inference engine that powers local AI model execution. That changed on July 7, 2026, when the mainline llama.cpp repository merged V4 architecture support in pull request 24162, followed by a critical fix for quantized key-value caches on the same day.

The KV cache fix matters more than it sounds. Before the patch, multi-turn conversations in every quantized version of V4-Flash were silently breaking. A tool-calling test tells the story: only 4 out of 15 function calls worked before the fix; after it, all 15 succeeded, matching DeepSeek's official implementation.

On July 31, 2026, DeepSeek released the official agent-tuned weights, meaning the model files available for download today are the production-ready version, not an interim preview. This convergence of software maturity and official release timing marks the moment V4-Flash transitioned from experimental to practical for local deployment.

Why Does "13 Billion Active Parameters" Not Mean What You Think?

DeepSeek V4-Flash uses a mixture-of-experts architecture, a design pattern where the model contains many specialized sub-networks and activates only a subset for each task. The model has 284 billion total parameters but only 13 billion active per token, which is why DeepSeek can serve it cheaply on their cloud API. However, this creates a critical misconception: V4-Flash does not behave like a 13-billion-parameter model on your hardware.

Here is the distinction that matters. Active parameters determine the compute cost per token, which is why the model is efficient to run remotely. Total parameters determine the memory bill. Because the router network can call any expert on any token, every expert must be resident in memory simultaneously. A typical 13-billion-parameter dense model needs about 8GB at 4-bit precision; V4-Flash at comparable precision needs roughly twenty times that amount.

How Much Memory Do You Actually Need?

The memory requirements depend on the quantization level, a process that compresses the model to fit on smaller machines. Unsloth, a specialist optimization team, published calibrated quantization files in July 2026 with measured quality at each tier, replacing the community conversions that dominated the fork era.

• 2-bit class (IQ2_XXS): 92 to 102GB file size with 110GB memory floor; achieves about 78 percent agreement with the full model on token choices, the lowest quality tier.
• 3-bit (UD-IQ3_XXS, recommended): 103GB file size with 110GB memory floor; Unsloth's recommended choice for 128GB machines, balancing quality and memory efficiency.
• 4-bit (UD-Q4_K_XL): 155GB file size with 162GB memory floor; achieves about 96 percent top-token agreement, near-lossless quality.
• 8-bit (UD-Q8_K_L): 162GB file size with 169GB memory floor; bit-identical to the official release, true lossless compression.

The most striking finding is that lossless quality costs almost nothing. V4-Flash was trained with quantization awareness, meaning its expert weights ship natively in MXFP4 format. Unsloth's 8-bit repack is bit-identical to the official release across all 1,328 tensors, yet only 7GB larger than the 4-bit version. If you have 169GB of memory, you are not running an approximation of V4-Flash; you are running V4-Flash exactly as DeepSeek released it.

The quality floor below 3 bits now has numbers attached. The 2-bit class agrees with the full model on about 78 percent of token choices, against roughly 96 percent at 4-bit. Below 3 bits, the difference is measurable and noticeable in practice, not speculation.

Which Hardware Can Actually Run This?

The practical tier map has shifted with mainline software support. A 32GB laptop or mini PC cannot run V4-Flash; the model simply does not fit. A gaming GPU with 24 to 32GB of VRAM can offload experts to system RAM but produces only single-digit tokens per second, making it impractical for real work.

The sweet spot is 128GB of unified memory, a configuration available in high-end Apple M4 Max laptops, AMD Ryzen AI Max+ 395 processors, and NVIDIA DGX Spark workstations. At this tier, the 3-bit quantization (UD-IQ3_XXS) fits comfortably at 103GB, leaving headroom for the operating system and runtime overhead. Reported speeds hover around 6 tokens per second at 3-bit on mainline llama.cpp, according to pull request 24162's published benchmarks.

For those with 192GB or more of unified memory, the 4-bit or lossless 8-bit builds become viable. June testing on a 192GB M3 Ultra Mac reported approximately 25 tokens per second, a speed that makes interactive use practical. Apple's new M5 Mac Studio, announced August 25, 2026, puts the 256GB tier back on sale starting September 22, 2026, expanding the comfortable tier further.

How to Run DeepSeek V4-Flash Locally Today

• llama.cpp command line: Download the model directly from Hugging Face using llama.cpp's built-in Hugging Face integration. One command pulls the quantized file and starts inference without any additional setup.
• Unsloth Studio: A graphical interface that handles model download, memory offload, and inference settings automatically. Search for DeepSeek-V4-Flash, select your quantization tier, and start chatting.
• LM Studio: A desktop application that added V4-Flash support in August 2026, ending its absence from graphical tools. The same announcement includes LM Studio Bionic, a hosted tier running on US-based servers with zero data retention by default.

One door that looks open is not: Ollama's library lists deepseek-v4-flash only as a cloud tag, meaning requests route to Ollama's servers rather than running locally. As of August 26, 2026, no native local entry could be confirmed in Ollama's offerings.

What Does This Mean for Self-Hosted AI?

The convergence of software maturity and hardware availability marks a turning point for local AI deployment. Six months ago, running a state-of-the-art reasoning model on your own machine required experimental forks and technical expertise. Today, it requires only standard software and 128GB of memory, a configuration that is expensive but no longer exotic.

The license remains MIT, unmodified, and the download is free. The bar for entry has moved from "experimental" to "practical," but the fundamental economics have not changed: you pay for hardware once, then run the model as many times as you want without per-token API fees. For teams processing sensitive data, running inference entirely on-premises, or optimizing for cost at scale, this shift unlocks a new tier of capability.

Beyond DeepSeek, the infrastructure supporting local models continues to mature. OpenCompany, an open-source alternative to Microsoft Copilot Studio, treats Ollama and LM Studio as first-class providers with automatic capability detection, meaning local models integrate seamlessly into multi-agent workflows without external API traffic. This ecosystem shift suggests that self-hosted inference is no longer a niche concern but an expected feature of modern AI infrastructure.