DeepSeek V4 Flash (284B MoE)

Positioning

DeepSeek V4 Flash is the late-2025 / early-2026 mid-tier MoE that takes the V4 Pro reasoning DNA and packages it for the homelab and consumer-tier hardware that V4 Pro can't comfortably target. 284B total parameters with 16-22B active per token (variant-dependent). Where V4 Pro asks for 192-GB unified memory or rented datacenter GPUs, V4 Flash fits a 128-GB Mac Studio M3 Ultra at Q4 (170 GB partial-offload, doable) or runs comfortably on the 192-GB tier. The operator-grade pitch: same DeepSeek-team training discipline + reasoning-first design as V4 Pro at one-fifth the active parameter count. For most readers of this site, V4 Flash is the right DeepSeek V4 to actually consider running.

Strengths

• Strong reasoning + coding combo at meaningfully lower hardware cost than V4 Pro. Benchmarks land 5-10 percentage points behind V4 Pro on most evals — which still beats most pre-V4-era frontier models.
• MoE efficiency. ~16-22B active parameters per token mean inference math is closer to a 22B dense model than a 284B dense one. Tok/s is reasonable for the absolute parameter count, especially on Apple Silicon where the Mac Studio M3 Ultra at 192 GB delivers ~18-28 tok/s on Q4.
• Permissive license — open weights, commercial use allowed (verify the DeepSeek license for your specific use case).
• Genuinely deployable on consumer + homelab hardware. ~170-180 GB at Q4 fits 192-GB tier; ~140 GB at Q3 fits 128-GB tier with offload. Multi-GPU homelab (4× RTX 3090 = 96 GB) runs Q3 with system-RAM offload. Real options.
• Strong day-zero tooling support. vLLM, SGLang, llama.cpp all shipped V4 Flash compatibility within hours of release. Less tooling lag than the V4 Pro path.

Limitations

• Quality gap vs V4 Pro is real, not just on-paper. For the hardest reasoning tasks (long multi-step proofs, edge-case code refactoring), V4 Pro is meaningfully better. V4 Flash is "frontier-adjacent," not "frontier."
• Memory is still substantial. ~170 GB at Q4 is far above any single consumer GPU. 24 GB cards are not in scope at any usable quant. The minimum operator-grade hardware is 96 GB unified memory or 4-card homelab.
• Tok/s drops fast at lower quants. Q3 at ~13-18 tok/s on Mac Studio M3 Ultra is functional but not interactive. Q2 quality regressions are visible vs Q3.
• Knowledge cutoff is early-2026 — for current-events / recent-API workloads, augment with RAG. Same constraint as V4 Pro.

Real-world performance on Mac Studio M3 Ultra (192 GB)

• Q4 (~170 GB): ~18-28 tok/s decode, TTFT ~1-2s on 1K prompts. Genuinely interactive.
• Q3 (~140 GB): ~22-32 tok/s, faster TTFT, slight quality dip. The right balance for daily use.
• Q5 (~200 GB partial-offload to swap): ~10-15 tok/s. Quality bump over Q4 is small; rarely worth the speed loss on local hardware.
• Compare with: rented H100 80GB ×4 datacenter setup runs FP8 V4 Flash at ~100-150 tok/s.

Should you run this locally?

Yes, if you have a 96-GB+ Mac Studio (or equivalent unified memory hardware) and want frontier-tier reasoning + coding output without the V4 Pro hardware cost. The Mac Studio M3 Ultra 192 GB tier at Q4 is genuinely interactive (~18-28 tok/s) — that's a usable daily driver, not just a batch-processing tool.

Yes, if you're running a 4-card homelab (RTX 3090 ×4 at $2,800-3,200 used) and willing to accept Q3 with partial offload. The economics work out — a Mac Studio M3 Ultra 192 GB is $5,000-7,000; the 4-card 3090 rig is half that.

No, for anyone running a single consumer GPU. 24 GB doesn't fit any usable quant. Use Qwen 3 30B-A3B or DeepSeek R1 Distill instead — designed for 24 GB tier hardware.

Probably not, for "I have a 192-GB Mac Studio and want absolute frontier" — pick V4 Pro at Q3 instead. Quality is meaningfully better; tok/s drops to 5-9 but it's the same hardware.

How it compares

• vs DeepSeek V4 Pro (1.6T MoE) → V4 Pro has higher quality ceiling at much higher hardware cost. V4 Flash fits 128-GB hardware where V4 Pro doesn't. For 95% of operators reading these verdicts, V4 Flash is the right DeepSeek V4 to actually deploy.
• vs Qwen 3.5 235B-A17B → Qwen 3.5 235B has Apache 2.0 license + better multilingual; V4 Flash has stronger reasoning chain-of-thought + slightly faster decode (16-22B active vs 17B active is essentially comparable). Pick on language requirements + license preference + reasoning-vs-generalist focus.
• vs Qwen 3 235B-A22B → similar hardware tier, similar quality tier. Q3 235B-A22B has Apache 2.0 license; V4 Flash has stronger reasoning. Both are reasonable picks; the choice often comes down to which the operator already has loaded.
• vs DeepSeek R1 (671B reasoning specialist) → R1 is reasoning-only; V4 Flash is generalist with reasoning. For reasoning-only workloads R1 wins; for mixed daily-driver tasks V4 Flash is more useful and runs on smaller hardware.
• vs Llama 4 Scout (Meta MoE) → Llama 4 Scout has 128k effective context vs V4 Flash's 64k. Llama license has 700M MAU clause; DeepSeek license is more permissive. Pick on context-length + license.

Run this yourself

# Mac Studio M3 Ultra 192GB — Q4 fits (partial offload tolerable)
ollama pull deepseek-v4-flash:q4_K_M
ollama run deepseek-v4-flash:q4_K_M

# Or via llama.cpp directly:
llama-server -m deepseek-v4-flash-Q4_K_M.gguf \
  --ctx-size 65536 -ngl 999 --no-mmap

Quant: Q4_K_M GGUF Context: 65536 (KV cache f16, ~30 GB additional) Backend: llama.cpp Metal via Ollama Hardware: Mac Studio M3 Ultra 192 GB unified memory