DeepSeek-V3 671B (MoE) KV-Cache Calculator
Per-token and total key-value cache memory for DeepSeek-V3 671B (MoE) across context length, batch size and cache precision.
DeepSeek-V3 671B (MoE): ~68.6 KB per token at FP16. DeepSeek-V3 is a 671B-parameter MoE with 37B active per token and Multi-head Latent Attention (MLA), which compresses the KV cache to a 512-dim latent per token — far below classic GQA.
Formula
Disclaimer: This tool is for general informational and estimation purposes only and is not professional financial, tax, accounting or legal advice. All figures are estimates — verify with a qualified professional before making decisions. Read the full disclaimer.
About DeepSeek-V3 671B (MoE) KV-Cache Calculator
The KV cache is the hidden memory cost of serving DeepSeek-V3 671B (MoE): every generated or prompted token stores its attention keys and values for reuse, and at long contexts this cache can rival the model weights themselves. This calculator uses DeepSeek-V3 671B (MoE)'s exact attention geometry — DeepSeek's Multi-head Latent Attention, which caches a single compressed 576-dim latent per token per layer — to give per-token, per-sequence and whole-batch cache sizes at FP16, FP8 and INT4 precision. Use it to size batch limits for your GPU or to see what a 128K-context request really costs.
How to use DeepSeek-V3 671B (MoE) KV-Cache Calculator
- 1Enter your values into DeepSeek-V3 671B (MoE) KV-Cache Calculator — sensible, domain-typical defaults are pre-filled so you see a real result immediately.
- 2The result recomputes live using the formula shown on the page; there is no button to press.
- 3Adjust any input to compare scenarios, then read the worked example to see the substituted numbers.
Why use DeepSeek-V3 671B (MoE) KV-Cache Calculator?
- ✓Computes DeepSeek-V3 671B (MoE) KV-Cache instantly in your browser — no sign-up, no upload, no server round-trip.
- ✓100% free and unlimited, with the exact formula shown: MLA cache/token = (512 latent + 64 rope) × layers × bytes = 576 × 61 × bytes (DeepSeek MLA compresses K,V into one lat.
- ✓Runs entirely client-side, so every value you enter stays private on your device.
- ✓Live recompute as you type, with a worked example and authoritative references for trust.
Frequently asked questions
How does MLA change the KV-cache math for DeepSeek-V3?+
Instead of caching full K and V per head, MLA caches one 512-dim compressed latent plus a 64-dim rotary key per token per layer: (512+64) × 61 layers × 2 bytes ≈ 70 KB per token at FP16 — comparable to a 7B GQA model despite 671B parameters.
What does it take to host DeepSeek-V3 weights?+
All 671B parameters must be resident: ~671 GB at FP8 (the native release format) or ~1.34 TB at BF16. That means a multi-node cluster (e.g. 2× 8×H200) regardless of the small active-parameter count, since the router may pick any expert.
Why does the KV cache matter more than weights for serving throughput?+
Weights are paid once per GPU; cache is paid per concurrent request and per token of context. Batch size — and therefore throughput — is capped by how many sequence caches fit in the VRAM left after weights, which is exactly what this tool computes.
What does paged attention change?+
PagedAttention (vLLM) allocates the cache in fixed-size blocks on demand instead of reserving the full context up front, eliminating fragmentation and letting you overcommit. The per-token cost shown here is unchanged — you just stop paying for unused reservation.
Related ML & AI tools
ROC-AUC Calculator (from TPR/FPR points)
Trapezoidal area under the ROC curve from your (FPR, TPR) operating points — the threshold-independent ranking score.
● LiveClassification Threshold Cost Calculator
Find the probability cutoff that minimizes expected cost given your false-positive and false-negative penalties.
● LiveSilhouette Score Calculator
Cluster cohesion vs separation for one point — the building block of the silhouette metric for choosing K.
● Live