Intermediate 14 minDeepSeek

DeepSeek V4 Flash 284B: the 1st frontier model that fits on a Mac Studio

DeepSeek V4 Flash 284B : the 1st frontier that fits on Mac Studio

DeepSeek V4 Flash 284B was released on April 24, 2026, at the same time as V4 Pro. It is the "efficient" variant: 284 billion total parameters in MoE (13B active per token), an MIT license, a 1M-token context, the same CSA+HCA architecture, and the same 3 thinking modes as Pro. Most importantly, it is the first frontier-class model that fits on a single workstation: 170 GB in Q4—a 256 GB Mac Studio M3 Ultra is enough. This guide explains how to install it and what it is worth.

By Mohamed Meguedmi·Update 2026-08-11·Tested on Windows, macOS, and Linux

#DeepSeek V4 Flash at a glance

Output
April 24, 2026, simultaneous with V4 Pro. Available on HuggingFace: deepseek-ai/DeepSeek-V4-Flash.
Architecture
MoE: 284B total, 13B active per token. 128 experts per layer, top-8 routing. Base + Instruct variants available.
Innovations inherited from the Pro
Caution: CSA+HCA, mHC, Muon optimizer, mixed FP4+FP8 training.
Context
1 000 000 native tokens (same as Pro).
Thinking modes
3 levels: No / High / Max — toggleable in the prompt.
License
MIT, same as Pro. No commercial or geographic usage restrictions.

#1. Why it’s an event

Before April 24, 2026, running a frontier model locally meant either a GPU cluster costing €80k+ or accepting the degraded quality of a 70B model. DeepSeek V4 Flash changes that: 170 GB in Q4 fits configurations costing €4k–8k used.

The complete kit for building your own AI assistant.

Benefit from our experience and a guided path to build your local assistant and use your AI every day.

  • Lifetime online access
  • PDF + files
  • Refunded within 30 days
Mac Studio M3 Ultra 256 GB
€7,299 listed through September 2026, now available secondhand (the M5 Ultra 256 GB replaces it at Apple). 170 GB model + 30 GB context = plenty of room. ~10–12 tok/s in Q4.
2× RTX 5090 32 GB (64 GB VRAM)
≈ €11,400. Insufficient—~110 GB is missing. Possible with CPU offload, but slow (3-5 tok/s).
4× RTX A6000 48 GB (192 GB VRAM)
~€12,000 used. Runs comfortably at ~18–25 tok/s.
Workstation: 2× Mac Studio Max 64 GB
Insufficient on a single system. With llama.cpp RPC swarm and 4× Mac Studio: possible, ~5 tok/s.
→
The real standout: Mac Studio Ultra
The M3 Ultra 256 GB (€7,299 at retail through September 2026, now available used) runs DeepSeek V4 Flash Q4 at ~10 tok/s in chat mode. No other personal machine in the world, at this price, runs a frontier model. Apple Silicon is what makes this possible (unified memory + 800 GB/s bandwidth).

#2. Architecture (efficient V4 variant)

V4 Flash is a proportional reduction of V4 Pro: ~5.7× fewer experts, ~3.8× fewer active parameters, but the same CSA+HCA + Muon + FP4/FP8 skeleton. This consistency lets V4 Flash inherit the Pro's quality at a much lower cost.

Experts per layer
128 (vs 256 for Pro). Identical Top-8 routing.
Active parameters
13B (vs. 49B for Pro). Inference speed ~3.8× higher at equivalent total params.
Layers
61 (Pro: 76). Moderate reduction to preserve reasoning depth.
Attention heads
Identical to the Pro (CSA+HCA configured identically). No savings on attention to preserve long-context quality.
Pretraining
“Efficient reasoning” target: 18T tokens (vs 32T for Pro), with a data mix optimized for math + code.

#3. Hardware by Quantization

Total VRAM required (model + 32k-token KV cache)
QuantizationModel VRAM+ KV 32kRecommended configurations
FP16 (reference)568 GB~620 GBDC: 4× H200 141GB. Impractical locally.
Q8_0305 GB~340 GBMac Studio M3 Ultra 512 GB or 8× RTX 4090 24GB.
Q5_K_M205 GB~235 GBMac Studio Ultra 256 GB (barely enough), 4× RTX A6000 48GB comfortable.
Q4_K_M170 GB~200 GBMac Studio Ultra 256 GB, 4× RTX 4090 24GB, 2× RTX 6000 Ada 48GB.
IQ3_M (compressed)130 GB~160 GBMac Studio M2 Ultra 192 GB, 4× RTX 4090 (light offload).
IQ2_XS (extreme)85 GB~115 GBRTX 5090 + 64 GB DDR5 offload, or 2× RTX 4090.
i
2026 sweet spot
The Mac Studio M3 Ultra 256 GB in Q4_K_M offers the best price/quality/speed ratio for serious personal use. €7,300 new through September 2026 (now available used, or as a new 256 GB M5 Ultra), ~10 tok/s, near-FP16 quality (Q4 loses <2% on benchmarks at this size). There is no equivalent on the PC side without exceeding €12,000.

#4. Mac Studio Ultra 256/512 GB setup

Complete Mac Studio M3 Ultra installation
# 1. Relever la mémoire GPU allouable
sudo sysctl iogpu.wired_limit_mb=240640    # 256 Go : 235 Go GPU
# Ou pour Mac Studio 512 Go :
# sudo sysctl iogpu.wired_limit_mb=491520

# Persistant
echo "iogpu.wired_limit_mb=240640" | sudo tee -a /etc/sysctl.conf

# 2. llama.cpp avec Metal + Flash Attention
git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp
make GGML_METAL=1 LLAMA_METAL_EMBED_LIBRARY=1

# 3. Téléchargement Q4_K_M (~170 Go split en chunks)
huggingface-cli download deepseek-ai/DeepSeek-V4-Flash-GGUF \
  --include "*Q4_K_M*.gguf" \
  --local-dir ./models

# 4. Lancement avec optimisations max
./build/bin/llama-server \
  -m ./models/DeepSeek-V4-Flash-Q4_K_M-00001-of-00004.gguf \
  -ngl 99 -fa \
  -ctk q8_0 -ctv q8_0 \
  -c 65536 \
  --host 0.0.0.0 --port 8080
→
Once started
The server exposes an OpenAI-compatible API at http://localhost:8080. Connect Open WebUI, Continue.dev, Aider, or Raycast AI to it—all will work with your local frontier model.

#5. NVIDIA multi-GPU setup

#Option A — vLLM 4× RTX A6000 (48 GB × 4 = 192 GB)

vLLM tensor parallelism
pip install vllm>=0.7

vllm serve deepseek-ai/DeepSeek-V4-Flash \
  --tensor-parallel-size 4 \
  --quantization fp8 \
  --max-model-len 1000000 \
  --enable-prefix-caching \
  --port 8000

#Option B — llama.cpp with partial offload

2× RTX 5090 + CPU offload
./build/bin/llama-server \
  -m ./models/DeepSeek-V4-Flash-Q4_K_M-00001-of-00004.gguf \
  -ngl 50 \
  --tensor-split 0.5,0.5 \
  -ctk q8_0 -ctv q8_0 \
  -c 32768
!
Offload = slow
Anything that does not fit in VRAM spills into system RAM (DDR5 at ~80 GB/s vs. HBM3 at 1 TB/s). With 50% offloaded, speed drops to 3-5 tok/s. For interactive use, aim for 100% in VRAM.

#6. Measured benchmarks and tokens/sec

DeepSeek-published benchmarks (April 24, 2026), Mode High values
BenchmarkV4 FlashV4 ProLlama 4 Scout 109BMistral Large 2
MMLU-Pro79.484.276.875.3
GPQA Diamond70.178.563.259.4
AIME 202576.587.061.352.1
LiveCodeBench v572.679.865.461.8
LongBench v2 (1M)68.972.6——
Measured tokens per second (chat, 4k context, Q4_K_M)
SetupTok/s mode NoTokens/s in High modeMax mode tok/s
Mac Studio M3 Ultra 256 GB10-128-105-7
Mac Studio M3 Ultra 512 GB12-1410-127-9
4× RTX A6000 48GB20-2516-2012-15
2× RTX 5090 + 64GB offload3-52-41-2

#7. Thinking modes and long context

V4 Flash inherits V4 Pro's three thinking modes, with slightly lower quality but much lower inference cost. It's the right choice for agents that need to reason without monopolizing infrastructure.

No Mode
Direct response. Very low latency (~1s to start streaming). For conversational chat, translation.
High mode
200–2000 chain-of-thought tokens before the answer. +25% quality on math/code. 5–15s latency before the first token of the final answer.
Max mode
Up to 16k reasoning tokens. 1-5 minutes of local latency. Reserve it for hard problems (math proofs, complex debugging).
1M long context
Same capabilities as V4 Pro thanks to HCA—Needle-in-Haystack 1M ~95% (vs. 98% for Pro). Excellent for RAG on large databases.

#V4 Flash vs. V4 Pro: choosing

Summary comparison
CriterionV4 Flash 284BV4 Pro 1.6T
Total parameters284 B1 600 B
Active parameters / token13 B49 B
Q4_K_M VRAM170 GB960 GB
Minimum local hardwareMac Studio Ultra 256 GB (used M3 Ultra or new M5 Ultra)~€80,000 cluster
Typical local tokens/sec10-250,5-2
MMLU-Pro quality79.484.2
AIME 2025 quality76.587.0
Native context1 M1 M
Max thinking modeYesYes (better quality)
LicenseMITMIT
→
When to choose Flash, when to choose Pro
V4 Flash: 95% of use cases—chat, code, RAG, agents, long context. The only frontier model that fits on a workstation. V4 Pro: pure research (math, proofs, competitive AIME), workflows where Max quality matters more than latency and hardware cost.

#Officialized on July 31, 2026 (build 0731)

DeepSeek released V4-Flash from preview in late July 2026 with build « 0731 »: the API is in public beta, and along the way the launch price of V4-Pro dropped by 75% ( $0.87 per million output tokens). Artificial Analysis places the V4 family "among the open-weight leaders." In terms of weights, nothing changes: 284B total with 13B active — still the only semi-practical V4 variant for high-end local use (256 GB Mac, Strix Halo, multi-GPU) with aggressive quantization.

Since August 2, LM Studio offers it too: locally if your machine can handle it, or through the Bionic app on U.S. servers with “Zero Data Retention” by default—a trade-off worth knowing about, even though this site's philosophy remains 100% local.

#FAQ

Does DeepSeek V4 Flash really run on a Mac Studio M3 Ultra 256 GB?+
Yes, in Q4_K_M: 170 GB model + 30 GB 32k context + 8 GB system = ~210 GB, within 256 GB. Measured speed: 10–12 tok/s in Non mode, 8–10 in High. This is the only single-machine Mac capable of this as of April 2026.
How much does a viable local setup for V4 Flash cost?+
Three budgets: (1) Mac Studio Ultra 256 GB — used M3 Ultra (€7,300 new through September 2026) or new M5 Ultra. (2) ~€12,000 — 4× used RTX A6000 48GB + workstation. (3) ≈ €11,400 — 2× RTX 5090 + 64 GB DDR5 + offload, but slow (3-5 tok/s).
Should you choose V4 Flash or wait for a 70B distillation?+
DeepSeek announced 70B and 32B distillations within 4-6 weeks. If you don't have a Mac Studio Ultra or a GPU cluster, waiting makes sense. If you already have one, V4 Flash now >> 70B distillation in 5 weeks (higher quality and 1M context).
Does V4 Flash beat Llama 4 Maverick / Scout?+
According to the benchmarks published by DeepSeek: yes, on all of them (MMLU-Pro 79.4 vs 76.8 for Scout, GPQA 70.1 vs 63.2). Independent confirmation is expected within 2 weeks via Chatbot Arena. V4 Flash's clear advantage: native 1M context, while Llama 4 remains at 256k.
How much power does a Mac Studio M3 Ultra consume during V4 Flash inference?+
Peak ~150 W during generation, ~30 W idle after loading. With 8h of active use: ~1.2 kWh = ~€0.30/day (FR 2026). An LED bulb uses more power in standby.
Is Max thinking mode worth it on V4 Flash?+
On Mac Studio: yes for difficult problems (math, proofs, debugging), but expect 1–5 minutes per response. For normal chat use, High mode is more than sufficient (8–10 tok/s, quality ~Mistral Large × 2).
Can you fine-tune V4 Flash locally?+
LoRA / QLoRA: yes, on a Mac Studio Ultra 512 GB via MLX-LM (support to be verified based on the installed version) or unsloth (NVIDIA, 4× A6000 minimum). Full fine-tuning: not practical locally — estimated training cost of $200–400k in the cloud.
Is there a stable Q4 quantized version as of April 25, 2026?+
Official GGUF Q4_K_M and Q5_K_M quants are coming out this weekend (26-27 April) on HuggingFace. AWQ quants for vLLM: already available (deepseek-ai/DeepSeek-V4-Flash-AWQ). Follow @DeepSeek_AI on X for announcements.

Prices change quickly: every Monday and Thursday, our tracker records the lowest price for local AI graphics cards, along with the price per GB of VRAM.

Did this guide help you?

Feedback, an error, or a clarification? Let us know—it improves the guide for everyone.

Prices in euros (€) are French market prices including VAT, as checked by BestLLMfor. US prices differ: the Amazon buttons show the current US price.