What it costs to run DeepSeek V4.1 Flash (MoE)
Trying to decide what to buy? The best GPU for DeepSeek V4.1 Flash (MoE) →
Memory needed, by quantisation
At 16k context. Weights + KV cache + ~0.8 GB overhead.
| Quant | Weights | KV cache | Total | Quality |
|---|---|---|---|---|
| F16 | 1421.6 GB | 0.0 GB | 1422.4 GB | Full precision. Reference quality, twice the size of Q8 for no practical gain in most chat use. |
| Q8_0 | 755.2 GB | 0.0 GB | 756.0 GB | Effectively lossless. Use when VRAM is not the constraint. |
| Q6_K | 582.8 GB | 0.0 GB | 583.6 GB | Very close to Q8 at meaningfully less memory. A safe high-quality pick. |
| Q5_K_M | 504.7 GB | 0.0 GB | 505.5 GB | Small, hard-to-notice quality loss. Good balance. |
| Q5_0 | 492.2 GB | 0.0 GB | 493.0 GB | Older-style 5-bit. Q5_K_M is usually the better pick at the same size. |
| Q4_K_M | 430.9 GB | 0.0 GB | 431.7 GB | The community default. Best quality-per-gigabyte for most people. |
| Q4_0 | 404.3 GB | 0.0 GB | 405.1 GB | Older-style 4-bit, measurably worse than Q4_K_M at a similar size. Avoid unless required. |
| IQ4_XS | 377.6 GB | 0.0 GB | 378.4 GB | Newer 4-bit, smaller than Q4_K_M with comparable quality. Needs a recent llama.cpp. |
| Q3_K_M | 347.4 GB | 0.0 GB | 348.2 GB | Noticeable degradation. Use to fit a larger model that would otherwise not run. |
| Q2_K | 297.6 GB | 0.0 GB | 298.4 GB | Heavy degradation. Almost always better to run a smaller model at Q4_K_M instead. |
| IQ3_XS | 293.2 GB | 0.0 GB | 294.0 GB | Aggressive. Usually better than a smaller model at Q4, but test before trusting it. |
Cards that can run it
Best quantisation each card fits at 16k context, with live rental price where we track one.
| GPU | VRAM | Best fit | Uses | ~tok/s | Rent from |
|---|---|---|---|---|---|
| Apple M3 Ultra | 512 GB | Q5_K_Mtight | 505 GB | — | — |
How these numbers are produced
- Model shape comes from the model's own config.json — 40 layers, 64 attention heads, 1 KV heads.
- Weight size is parameters × effective bits-per-weight. Those constants are checked against real published quantised file sizes — median error 0.7% across 40 measurements.
- KV cache is 2 × layers × KV-heads × head-dim × context × 2 bytes. Using KV-heads rather than attention heads is what makes this correct for grouped-query attention; treating a GQA model as multi-head overstates the cache by up to 8×.
- Tokens/sec is an estimate, not a benchmark. Generation is memory-bandwidth-bound, so this is bandwidth ÷ weight-bytes derated to 75%. Real throughput depends on your runtime, batch size and kernels. We don't run our own hardware tests — see our editorial policy.
- This is a mixture-of-experts model. Generation reads only the experts active for each token (6 of 384), so throughput is scored on ~16.0B active parameters, not the full 763B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Causal encoder-decoder (20 + 20 layers) with CSA2 sparse attention and FP4 main KV caching: the model card states the global KV cache is 890 bytes per token. Mapped so the engine's fp16 formula gives about that: 40 layers x 1 KV head x head_dim 6 x 2 x 2 B = 960 B/token (~1 GB at 1M context); the bounded sliding-window replay state is extra and small. head_dim 6 is a mapping, not the model's 512; a q8/q4 KV setting in the calculator would understate it. Active = 16B per token while decoding (8B during prefill), vendor figure. params = HF safetensors total (~763B) = 552B backbone + Engram conditional-memory tables (196B, token lookups) + vision encoder, all of which have to be stored somewhere. HF also lists a 1017 revision; Novita serves both at the same price. HF config.json + model card read 29 Oct 2026.
- Some memory-bandwidth figures are not yet independently verified.Where that's the case we show no tokens/sec at all rather than a number we can't stand behind.
- Rental prices are pulled hourly from provider APIs (last updated Thu, 29 Oct 2026 04:20:06 GMT). Spot/interruptible pricing can change or vanish without notice.
Try it interactively
Drag the context slider and watch the KV cache fill the card
Worth reading before you buy
GLM-4.5-Air on local hardware
The consumer-runnable GLM: VRAM by quant, which cards and unified-memory boxes clear it, and what to expect from published numbers — the written companion to this page.
Cheapest way to run GLM in the cloud
Air on one 80 GB card, GLM-4.6 on two, GLM-5 on a pair of H200s — the multi-GPU maths with live, dated rental prices.
Run Mistral Small 24B locally
VRAM by quant (Q4 ~14 GB, Q8 ~25 GB, BF16 ~55 GB), which cards clear it, what tokens/s to expect from published benchmarks — and what a 24 GB card rents for when yours doesn't.
Cheapest RTX 4090 rental
Live per-hour 4090 prices on Vast and Runpod, the spot-vs-on-demand lever, and what a 24 GB card actually runs — the rent-a-card answer for everything in the 14B–32B class.