DeepSeek-V4 Pro

1.6 trillion parameters with 49B active. Nothing in this catalogue can hold it at 4-bit — it is here specifically to show where single-box local inference stops and a cluster begins. Its MLA attention keeps the KV cache small even at a million tokens, which is the one thing that does work in its favour.

DeepSeekMixture of expertsreasoningreasoning1 machines can run it

Run locally →

Total parameters

1.6T

Official spec sets memory need

Active parameters

49B

Official spec sets decode speed

Context

Official spec tokens

4-bit weights

871 GB

Calculated before KV cache

This is a sparse mixture of experts. All 1.6T of weights must be available to the runtime, but only 49B are read per generated token. Ordinary runtimes keep the full quantized model in memory; a runtime-specific SSD-streaming artifact can retain a smaller working set and fetch expert data on demand, trading speed for capacity.

Model specification

Publisher

DeepSeek Official spec

Architecture

Sparse mixture of experts Official spec

Total parameters

1.6T Official spec

Active parameters

49B per token Official spec

Context length

1,048,576 tokens Official spec

Attention ⓘ

MQA · 1 KV heads × 512 dim × 61 layers Official spec

Licence

MIT Official spec source

Specification confidence ⓘ

High · verified 2026-11-02 Official spec source

Released

1 May 2026 Official spec

Official source

Hugging Face

Quantizations and memory

Quantization	Format	Bits/weight	Resident	Download	Quality kept
`MLX 4-bit`	mlx	4.5	871.1 GB	871.1 GB	98.0%
`Q4_K_M`default	gguf	4.85	929.8 GB	929.8 GB	98.5%
`Q5_K_M`	gguf	5.69	1090.8 GB	1090.8 GB	99.3%
`Q8_0`	gguf	8.5	1613.7 GB	1613.7 GB	99.9%

Generic weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead. Runtime-specific artifacts use their published resident and download footprints; streamed models can therefore require much more disk than memory. Quality retention is an assumption, not a measured evaluation.

Coding & quality benchmarksCompare coding results →

No published benchmark results have been imported for this model yet. This is missing evidence, not a score of zero.

We show source metrics rather than deriving one opaque quality number. Different benchmarks measure genuinely different things, and collapsing them into a single score would hide exactly the disagreements worth seeing.

Recommendations

Cheapest that can run it

NVIDIA DGX H200 (8x H200, 1,128 GB)

~306–441 t/s · $359,430

Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.

Cheapest above 20 t/s

NVIDIA DGX H200 (8x H200, 1,128 GB)

~306–441 t/s · $359,430

20 tokens/second is roughly the point at which generation keeps pace with reading. Below it, interactive use feels like waiting.

Cheapest above 40 t/s

NVIDIA DGX H200 (8x H200, 1,128 GB)

~306–441 t/s · $359,430

40 tokens/second is the threshold most people describe as comfortable for coding agents, where output arrives faster than you can review it.

Fastest with real measurements

Nothing in the database qualifies.

Highest throughput among configurations with an actual published measurement rather than our estimate.

Best throughput per purchase-price unit

NVIDIA DGX H200 (8x H200, 1,128 GB)

~306–441 t/s · $359,430

Highest decode tokens/second per 1,000 units of the displayed purchase currency. Ignores running costs and resale — see the economics section for the full picture.

Hardware that runs DeepSeek-V4 Pro0 measured, 1 estimated

Quantization

Minimum decode: 0 t/s

1 of 1 rows

Hardware↕	Quant	Memory↕	Decode▼	Prefill↕	Context	Price↕	Fit	Confidence↕
NVIDIA DGX H200 (8x H200, 1,128 GB) NVIDIA · 1,128 GB · 38,400 GB/s	`Q4_K_M` recommended	962.5 GB	~306–441 t/s	~9560–19850 t/s	8K	$359,430	Borderline	Estimated

Submit a benchmarkContributions are reviewed before publication.

Hosted alternatives for this exact model

These endpoints serve the same open weights, so comparing them against local ownership is a like-for-like economic question rather than a quality trade-off.