Run DeepSeek V4-Flash Locally: Hardware Requirements 2026

Last updated: August 2026

Yes — with a 110GB-plus memory budget and mainline software, DeepSeek V4-Flash now runs locally, no forks required. That answer changed on July 7, 2026, when llama.cpp's V4 support matured and Unsloth shipped calibrated builds. In June, this page told you to wait for exactly those two events. Both happened.

This is the update the original article promised: memory math rebuilt on measured files, a tier table for the mainline era, and June's two watch signals graded on the record — along with the one we got half right. The license is still MIT, the download is still free, and the bar just moved.

What V4-Flash Actually Is, and the 13B Misconception

DeepSeek released the V4 family on April 23, 2026 as a two-model preview: V4-Pro at 1.6 trillion total parameters and V4-Flash at 284 billion. Flash is the one that matters at home; its official repository — 3.1 million downloads in the past month — explains both the excitement and the problem.

The misconception to kill immediately: "13 billion active" does not mean V4-Flash behaves like a 13B model on your hardware. Active parameters set the compute per token, which is why DeepSeek can serve it cheaply. Total parameters set the memory bill: the router can call any expert on any token, so every expert must be resident. A 13B dense model needs about 8GB at 4-bit; V4-Flash at comparable precision needs roughly twenty times that.

The capability is genuine: DeepSeek reports 79.0 percent on SWE-bench Verified, and Artificial Analysis placed Flash within a few points of the far larger V4-Pro on its intelligence index. The license is the real, unmodified MIT text. The legal side of self-hosting was solved in April; as of July, the software side is too.

The Memory Math, Now With Receipts

Quantization is still the only lever that shrinks the memory bill, but the files worth planning around changed: Unsloth's calibrated GGUF ladder replaced the fork-pinned community conversions, with measured quality at every rung, checked tensor-by-tensor against the official weights.

Sizes, floors, and quality from Unsloth's published measurements, July 2026. Unsloth also lists a 92GB floor for a 1-bit build; the official checkpoint is 156.4GB. June's "81GB" was this same 2-bit class in different units.

Two findings deserve the headline. First, lossless costs almost nothing: V4-Flash was quantization-aware-trained, its expert weights ship natively in MXFP4, and Unsloth's 8-bit repack is bit-identical to the release build across all 1,328 tensors by its published check — only 7GB larger than the 4-bit. Hold 169GB and you are not running an approximation of V4-Flash; you are running V4-Flash. Second, June's quality-floor warning now has numbers: 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, you feel it — measured, not vibes.

The KV cache footnote still applies: the million-token context window is priced for datacenters. At home, plan on 8K to 32K tokens and budget the gap between each file and its floor above for cache, runtime, and operating system.

The Software Gap Closed in July

In June, this section documented a gap: no stable release of llama.cpp, Ollama, or LM Studio could load the V4 architecture, and every consumer run depended on experimental forks. That era ended in two commits. Mainline llama.cpp added V4 support in pull request 24162, and a July 7 follow-up fixed quantized KV caches — the bug quietly wrecking multi-turn conversations in every provider's GGUF files. Unsloth's tool-calling test tells the story in one line: 4 of 15 before the fix, 15 of 15 after, matching DeepSeek's official code.

The fork era exits with its footnote earned. Salvatore Sanfilippo's experimental build carried this model through spring, and its quantizations are now measured rows in Unsloth's benchmark tables — the fork proves the demand, the mainline absorbs the code, the specialist industrializes the quality.

Running it today takes one of three doors, none involving anyone's personal runtime. llama.cpp pulls and serves the model directly, one command, ollama-style:

./llama.cpp/llama-cli \
    -hf unsloth/DeepSeek-V4-Flash-0731-GGUF:UD-IQ3_XXS \
    --temp 1.0 --top-p 1.0

Or use Unsloth Studio, the open-source local UI that handles the download, memory offload, and settings for you:

curl -fsSL https://unsloth.ai/install.sh | sh
unsloth studio

Then search for DeepSeek-V4-Flash and pick the quant your memory clears. Reasoning controls now work natively in llama.cpp too; Unsloth's guide carries the exact flags and stays fresher than any article snapshot.

The third door opened in August: LM Studio now lists V4 Flash for local download, ending its absence from the GUI apps. The same announcement adds a hosted tier, LM Studio Bionic, running the model on US-based servers with zero data retention by default — worth knowing as a jurisdiction alternative to DeepSeek's first-party API, though hosted is hosted, whatever the flag. One door still looks open and is not: Ollama's library lists deepseek-v4-flash only as a cloud tag — requests route to Ollama's servers, the same cloud-suffix pattern we flagged on Kimi K3's listing, with no native local entry we could confirm as of August 26. For genuinely local runs, the doors are llama.cpp, LM Studio, and Unsloth Studio.

Tier by Tier: Can Your Machine Run It Now?

The tier map, rebuilt for the mainline era. Speeds remain the softest numbers here, so each carries its source and date.

Speed sources: 96GB tier, June fork-era testing; GB10 figure, pull request 24162's published runs; 192GB tier, June MLX reports. Ballparks.

The story is the middle rows. In June, 128GB unified memory was "the true borderline": 2-bit quality on experimental software. Today it is the sweet spot — the recommended 103GB build, at a quality tier well above the old 2-bit compromise, on stock tooling. M4 Max Macs, AMD's Ryzen AI Max+ 395 boxes, and NVIDIA's DGX Spark sit in this class, and they are the machines you can actually buy.

The top row's supply twist resolved on August 25. Apple announced the M5-generation Mac Studio: the M5 Ultra starts at 96GB of unified memory and configures to 256GB for $4,000 more, shipping September 22, with a 512GB option arriving in late October, per Apple's published specifications. The M5 Max model tops out at 128GB from $2,499 — a first-party entrant in the sweet-spot tier. The comfortable seat is purchasable again; whether it is worth the memory step is its own question, and our M5 Ultra wait-or-buy breakdown runs that math.

The 96GB workstation tier still runs the model, its tradeoff now quantified: fully-on-GPU means the 2-bit class, and that means the measured 78 percent tier. Possible remains the right word; sensible begins at 128GB.

The RAM Crisis Still Owns the Escape Hatch

The consolation path for MoE models — attention on the GPU, experts in system RAM — still works mechanically and still fails economically. Tom's Hardware's RAM price index, updated July 1, lists the cheapest 32GB DDR5 kits around $369 and a 128GB kit near $1,834 — the capacity an offload build wants costs more than hardware that runs whole models in VRAM. The full picture lives in our RAM shortage explainer.

V4-Flash vs Kimi K3: What the Two Open Frontiers Ask of You

2026's two headline open releases bracket the local AI question from opposite ends.

Kimi K3 is the frontier you can license but not house: even its 1-bit quantization wants a 650GB-plus floor — two networked machines or a retired server, per our Kimi K3 reality check. V4-Flash is the frontier that fits: one 128GB computer, mainline software, unmodified MIT. On which open model changes what you can do at home this year, it is not close.

The Other 128GB Frontier: GLM-5.3-Flash

As of August 26, V4-Flash has a same-weight-class rival: Z.ai's GLM-5.3-Flash, released that morning under the same unmodified MIT license, aimed at the same 128GB machines. Here is where the matchup stands on day one.

GLM-5.3-Flash column reflects launch-day status, August 26, 2026. Its sizes are lineage-scaled estimates pending upload completion.

The day-one verdict is not close, and it is not permanent. V4-Flash 0731 holds the 128GB crown on the three things that matter for a machine you rely on: mainline software, a measured quality ladder, and agent-tuned official weights. GLM-5.3-Flash is the credible challenger — similar memory math, a bigger family, and the one card V4-Flash cannot play, native image and video input. Our GLM-5.3-Flash reality check carries its full hardware breakdown. When mainline llama.cpp lands glm5_next support, this table gets a rematch.

The 0731 Asterisk, Retired

July's caveat is closed. When DeepSeek announced the re-post-trained 0731 build on July 31, this page warned that the agent-tuned model lived only behind the API. The open weights landed on Hugging Face the same day — 166.9GB, MIT, superseding the April preview — and Unsloth's 0731 GGUF ladder followed with the same landmark numbers: lossless Q8 at 162GB, only 7GB over Q4. The checkpoint ships with the DSpark speculative-decoding module attached — the repository totals 304B parameters including that draft module, the base architecture unchanged — and Unsloth reports up to 2x faster decoding with it enabled. Local and hosted V4-Flash are the same model again, and every build this page recommends now points at the 0731 ladder. Our breakdown of the 0731 release covers what the re-post-training changed.

What to Run Below 110GB

If your machine sits under the floor, the right move is unchanged: run what fits well rather than what barely loads. On 16GB to 32GB machines, the small-MoE class — Gemma 4's 26B mixture among them — runs at conversational speed with no GPU; our mini PC guide for local AI maps those tiers, and what AI models actually run at home sorts the field by memory. With a 24GB card, 30B-class models run in VRAM and 70B-class models run quantized; the renewed RTX 3090 remains the best value per gigabyte of VRAM, as our local AI hardware guide covers.

Check Price on Amazon: NVIDIA RTX 3090 24GB (Renewed)

At 64GB of unified memory, the strongest small MoE models run with full working context on a box that doubles as a home server — and machines shipping with memory installed sidestepped the worst of the RAM market.

Check Price on Amazon: MINISFORUM AI X1 Pro 370 (64GB)

The hosted route stays fast and cheap, with the tradeoff this site exists to flag: your prompts are processed under the provider's policy and jurisdiction — for DeepSeek's first-party API, servers under Chinese jurisdiction. Our launch coverage walks the alternatives. Open weights keep the local option permanently yours, and as of July, far more readers can finally exercise it.

The Signal Ledger: Graded and Reposted

In June this article named two events that would move V4-Flash from expert project to weekend project. Both fired within four weeks: the mainline merge (pull request 24162, matured by the July 7 KV fix) and Unsloth's turnkey calibrated builds, lossless flag included. The half-right prediction: "Ollama and LM Studio support typically follows within weeks." llama.cpp itself became the one-command path first — and LM Studio's native listing did arrive, in August.

July posted three more signals. Graded as of August 26: the 0731 weights fired the same day they were posted, July 31, flipping local runs to the agent-tuned build. The native-listings signal half-fired — LM Studio local is live, while Ollama still shows only the cloud tag. And the Mac Studio ceiling resolved on August 25 with the M5 generation: the comfortable tier returns to sale September 22, at a $4,000 step from 96GB to 256GB. Two signals now stand posted for the next update: a native Ollama listing, which would make this a true appliance for the one-click crowd, and mainline glm5_next support, which arms the challenger in the table above. Dated updates as each lands.

The closing observation stands, upgraded: 13 billion active parameters at near-frontier quality is now a model an individual can download, verify bit-for-bit, and run on a machine they can buy. Whoever controls the infrastructure controls the experience — and this one's infrastructure quietly finished being built.

Frequently Asked Questions

Can an RTX 4090 or RTX 5090 run DeepSeek V4-Flash?

Not practically. The lightest builds want a 92GB-plus floor of combined RAM and VRAM, far beyond 24 or 32GB of VRAM, and offloading experts to system RAM stays single-digit slow in a year when that much RAM costs more than the card. Spend these GPUs on 30B and 70B-class models instead.

Does Ollama or LM Studio support V4-Flash?

LM Studio, yes — as of August it lists V4 Flash for genuine local download, per LM Studio's announcement, alongside a hosted Bionic tier on US-based servers with zero data retention by default. Ollama still lists deepseek-v4-flash only as a cloud tag as of August 26, routing requests to its servers. The fully local doors are llama.cpp, LM Studio, and Unsloth Studio.

How much memory does the 4-bit or lossless version need?

The 4-bit UD-Q4_K_XL is a 155GB file wanting about 162GB of combined memory; the lossless 8-bit is 162GB with a 169GB floor, per Unsloth. The 7GB between them is the whole cost of running the exact release checkpoint — 192GB-class machines are full-quality territory.

Will a 128GB MacBook Pro or Mac Studio run it?

Yes — this tier moved from borderline to recommended. The 103GB 3-bit build is Unsloth's pick for 128GB devices and loads through mainline llama.cpp, LM Studio, or Unsloth Studio. Expect a roughly 110GB working floor, modest speeds, and context in the 8K to 32K range.

Is the MIT license really safe for commercial use?

Yes. V4-Flash ships under the standard, unmodified MIT license, which permits commercial use, modification, and fine-tuning. Not every 2026 open-weight release works this way: MiniMax M2.7, for example, uses a modified MIT license with commercial-use conditions that deserve a careful read before any deployment.

Are the downloadable weights the same as the API model now?

Yes, since July 31. DeepSeek published the official 0731 checkpoint on Hugging Face the same day its API moved to that build, so local and hosted V4-Flash run the same agent-tuned model. The 0731 release supersedes the April preview and ships with the DSpark speculative-decoding module attached, which Unsloth supports for up to 2x faster local decoding.

Is DeepSeek V4-Flash or GLM-5.3-Flash better on a 128GB machine?

Today, V4-Flash: it has mainline llama.cpp and LM Studio support, a measured quality ladder, and agent-tuned official weights, while GLM-5.3-Flash launched August 26 with fork-only tooling and estimated sizes. GLM-5.3-Flash counters with native image and video input under the same MIT license. Revisit the comparison once mainline support lands for its architecture.