Can Titan V 12GB run DeepSeek-V4-Flash?

NVIDIA · 12 GB HBM2 · 652.8 GB/s bandwidth · DeepSeek · 158.1B (active 13B) · MoE · model ctx up to 1000K

CUDA

No — DeepSeek-V4-Flash at Q4 needs ≈ 88.12 GB even at 4K context, beyond the Titan V 12GB's 12 GB HBM2. It would only run with heavy CPU/disk offload. FP16 needs ≈ 317.52 GB — does not fit on this card. Estimates come from memory-bandwidth math; rows tagged measured override estimates. Relative ranking is more reliable than absolute tok/s. Methodology.

Fit & speed by quant and context

Quant	Context	VRAM needed	Fits?	tok/s (decode)
Q4	4K	88.12 GB	No	—
Q4	8K default	88.28 GB	No	—
Q4	32K	89.25 GB	No	—
FP16	4K	317.36 GB	No	—
FP16	8K default	317.52 GB	No	—
FP16	32K	318.5 GB	No	—

Fits = weights + KV(ctx) + 1 GB overhead ≤ 95% of VRAM. Measured anchors are context-agnostic; the fit verdict is recomputed per context. A missing tok/s means the model is far beyond this card (offload-only territory).

VRAM breakdown at 8K context

Quant	Weights	KV cache	Overhead	Total needed	Titan V 12GB VRAM
Q4	86.95 GB	0.33 GB	1 GB	88.28 GB	12 GB
FP16	316.2 GB	0.33 GB	1 GB	317.52 GB	12 GB

Other GPUs that run DeepSeek-V4-Flash

• B200 192GB SXM — ≈391.6 tok/s Q4 est.
• MI300X 192GB — ≈259.4 tok/s Q4 est.
• H200 141GB SXM5 — ≈235.0 tok/s Q4 est.
• Gaudi 3 128GB — ≈160.4 tok/s Q4 est.
• Mac Studio M3 Ultra 96GB — ≈40.1 tok/s Q4 est.
• Mac Studio M2 Ultra 192GB — ≈39.2 tok/s Q4 est.
• MacBook Pro M4 Max 128GB — ≈26.7 tok/s Q4 est.
• DGX Spark (GH200 480GB) — ≈24.5 tok/s Q4 est.

All GPUs for DeepSeek-V4-Flash →

Other models for the Titan V 12GB

• MiniCPM5-1B — ≈384.6 tok/s Q4 est.
• Qwen3.5-2B — ≈207.7 tok/s Q4 est.
• DeepSeek-R1-0528-Qwen3-8B-layer-mix-bpw-3.8-mlx — ≈180.6 tok/s Q4 est.
• Llama-3.2-3B-Instruct — ≈129.8 tok/s Q4 est.
• Qwen3-4B-Instruct-2507 — ≈103.9 tok/s Q4 est.
• NVIDIA-Nemotron-3-Nano-4B-BF16 — ≈103.9 tok/s Q4 est.
• Qwen3-4B-Thinking-2507 — ≈103.9 tok/s Q4 est.
• gemma-4-E2B-it — ≈81.5 tok/s Q4 est.

All models on the Titan V 12GB →