Open Weights Explained: Running DeepSeek V4 Flash

DeepSeek published V4-Flash-0731 on July 31, 2026 under an MIT license, and Artificial Analysis scored it 50 on its Intelligence Index, well above the median of 25 across the 101 same-class models it is compared against. On Artificial Analysis's intelligence-versus-cost chart, the vendor legend runs DeepSeek, Google, Anthropic, Z AI, OpenAI, SpaceXAI, Kimi, NVIDIA, MiniMax, Meta, Xiaomi, Tencent and Thinking Machines, with proprietary and open-weight releases plotted on the same axes.

Start with the hosted API at 14 cents a million

Open weights means the trained file is published and you are allowed to take it. It does not mean you have to be the one running it, and for most people starting out, the hosted version is the sane entry point.

DeepSeek's own API charges $0.14 per million input tokens and $0.28 per million output tokens, against a $0.43 and $1.20 median for comparable models on Artificial Analysis's numbers. Cached input is the outlier: $0.003 per million tokens, a 98% discount, which Artificial Analysis ranks first out of 101 models. Context window is 1M tokens. Text in, text out, no images.

A million tokens of output costs 28 cents. Hold onto that number, because it reframes what the download is for. Self-hosting an open model in 2026 is a decision about control, privacy and vendor dependency. The bill is no longer the argument for it.

One caveat sits in the same source. Artificial Analysis lists the model as available through exactly one API provider. Whatever else the MIT license permits, the convenient path today still runs through DeepSeek's servers.

162GB is the real gate

V4-Flash-0731 has 284 billion total parameters and 13 billion active. It is a Mixture-of-Experts model, meaning only a fraction of the network fires for any given token. Artificial Analysis lists those two figures on the same page as the $0.14 input price without connecting them, and DeepSeek has not published an explanation of how it set that price.

What decides whether you can run the model at home is the file size of the quantization, a compressed version of the weights that trades a little precision for a lot of disk and memory. Unsloth's GGUF repository publishes a full ladder of them. Per Unsloth's model card, Q8 (UD-Q8_K_XL) is the "full precision lossless" option at 162GB, and it is only 7GB bigger than the Q4 build.

162GB of weights wants a workstation with serious memory, and that figure is the honest gate on the phrase "run it locally." It applies to this model specifically rather than to open weights generally. The card also reports 48,707 downloads in the last month, so the hardware wall is clearly not stopping everyone.

The license is the other half of what you are downloading. MIT permits commercial use, which is the difference between a model you can build a business on and one you can only experiment with.

Load a GGUF file in Unsloth Studio

Nothing about open weights requires a command line here. Unsloth's card states you can run V4-Flash-0731 in Unsloth Studio with toggles for High and Max thinking, and it points readers at its own "How to Run DeepSeek-V4-0731" guide plus a separate DeepSeek-V4 guide for quantization analysis and instructions. Those are the vendor's own routes, and the card names no other application.

Pick the rung of the ladder before you start any download. The repository lists thirteen builds: UD-IQ1_S and UD-IQ1_M at the aggressive end, then UD-IQ2_XXS, UD-IQ2_M, UD-IQ3_XXS, UD-IQ3_S, UD-IQ4_XS, UD-IQ4_NL, UD-Q2_K_XL, UD-Q3_K_M, UD-Q3_K_XL, UD-Q4_K_XL and UD-Q8_K_XL. The card publishes sizes only for the top of that ladder, the 162GB Q8 and the Q4 sitting 7GB below it. What the smaller rungs weigh is not stated, though the file counts sketch the shape: the IQ1, IQ2 and Q2 builds ship in three parts, the IQ3, IQ4 and Q3 builds in four, and the Q4 and Q8 builds in five.

NL2Repo is 54.2 against Opus-4.8's 69.7

DeepSeek's own model card publishes a benchmark table comparing V4-Flash-0731 to its preview version, to DeepSeek-V4-Pro (Preview), to Z AI's GLM-5.2 and to Anthropic's Opus-4.8. The relevant comparison for anyone choosing an open model is the middle column.

Read down the columns and a pattern shows up. On terminal use and agent exams the open models are within a couple of points of the closed frontier, and on repository-scale code generation they are not: 54.2 against 69.7 on NL2Repo is a wide gap that no pricing argument closes.

Treat the table as a vendor claim, because it is one. DeepSeek ran these with its own harness, which the card says is still "to be released," at max reasoning effort with temperature 1.0 and top_p 0.95. Two of the rows on the full card, DSBench-FullStack and DSBench-Hard, are DeepSeek's internal test sets, which nobody outside the company can reproduce.

Read the 50 as a score for the hosted model

Every independent number in this post describes the hosted model at maximum reasoning effort. Artificial Analysis spent $72.02 running its nine evaluations, and the model emitted 210M output tokens doing it, more than double the 100M median for its class. Its speed row reads N/A, so there is no third-party measurement of how fast this model actually generates.

The file you download is a different artifact. Quantization to IQ2 or IQ1 changes the weights, and Unsloth's claim that its Dynamic 2.0 quants achieve "superior accuracy" over other quants is the vendor's own, published alongside the vendor's own quantization analysis. Artificial Analysis maintains an Openness Index that scores how open a model's release is, and an Intelligence Index that scores hosted APIs. Both stop at the release DeepSeek serves.

The International AI Safety Report 2026, chaired by Yoshua Bengio with input from over 100 experts, names a version of this problem at the frontier: an "evaluation gap" where "performance on pre-deployment tests does not reliably predict real-world utility or risk." The same gap runs straight through the consumer end of open weights, where the released file and the evaluated system are not the same thing.

Anyone can download 284 billion parameters under an MIT license, 48,707 people did it last month, and the score attached to those weights everywhere they are discussed was measured on DeepSeek's hosted API at max reasoning effort. The independent index covers that configuration. The compressed files people run at home come with the compressor's own grade on them, and the slot sitting open in this field is an outside score for the rung you actually downloaded.