ALITEQ.

What it costs to run Mixtral 8x7B (MoE)

46.7B paramsMoE · 8 experts, 2 active32k max contextapache-2.0

Memory needed, by quantisation

At 16k context. Weights + KV cache + ~0.8 GB overhead.

QuantWeightsKV cacheTotalQuality
F1687.0 GB2.0 GB89.8 GBFull precision. Reference quality, twice the size of Q8 for no practical gain in most chat use.
Q8_046.2 GB2.0 GB49.0 GBEffectively lossless. Use when VRAM is not the constraint.
Q6_K35.7 GB2.0 GB38.4 GBVery close to Q8 at meaningfully less memory. A safe high-quality pick.
Q5_K_M30.9 GB2.0 GB33.7 GBSmall, hard-to-notice quality loss. Good balance.
Q5_030.1 GB2.0 GB32.9 GBOlder-style 5-bit. Q5_K_M is usually the better pick at the same size.
Q4_K_M26.4 GB2.0 GB29.2 GBThe community default. Best quality-per-gigabyte for most people.
Q4_024.7 GB2.0 GB27.5 GBOlder-style 4-bit, measurably worse than Q4_K_M at a similar size. Avoid unless required.
IQ4_XS23.1 GB2.0 GB25.9 GBNewer 4-bit, smaller than Q4_K_M with comparable quality. Needs a recent llama.cpp.
Q3_K_M21.3 GB2.0 GB24.0 GBNoticeable degradation. Use to fit a larger model that would otherwise not run.
Q2_K18.2 GB2.0 GB21.0 GBHeavy degradation. Almost always better to run a smaller model at Q4_K_M instead.
IQ3_XS17.9 GB2.0 GB20.7 GBAggressive. 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.

GPUVRAMBest fitUses~tok/sRent from
Apple M3 Ultra512 GBF1690 GB~24
Apple M2 Ultra192 GBF1690 GB~24
Apple M4 Max (16-core CPU / 40-core GPU)128 GBF1690 GB~16
AMD Ryzen AI Max+ 395 (Strix Halo, 128 GB)128 GBF1690 GB~7
NVIDIA DGX Spark (GB10 Grace Blackwell)128 GBF1690 GB~8
NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition96 GBF16tight90 GB~52$1.690/hr
NVIDIA RTX PRO 6000 Blackwell Workstation Edition96 GBF16tight90 GB~52
Apple M2 Max96 GBF16tight90 GB~12
Apple M4 Pro64 GBQ8_049 GB~15
NVIDIA RTX 6000 Ada Generation48 GBQ6_K38 GB~68$0.428/hr
NVIDIA RTX A600048 GBQ6_K38 GB~55$0.330/hr
AMD Radeon PRO W790048 GBQ6_K38 GB~61
Apple M4 Max (14-core CPU / 32-core GPU)36 GBQ5_K_Mtight34 GB~34
NVIDIA RTX 5000 Ada Generation32 GBQ4_K_Mtight29 GB~55$0.490/hr
NVIDIA GeForce RTX 509032 GBQ4_K_Mtight29 GB~172$0.058/hr
NVIDIA RTX A500024 GBQ2_K21 GB~107$0.160/hr
NVIDIA GeForce RTX 409024 GBQ2_K21 GB~140$0.136/hr
NVIDIA GeForce RTX 3090 Ti24 GBQ2_K21 GB~140$0.116/hr
AMD Radeon RX 7900 XTX24 GBQ2_K21 GB~134
NVIDIA GeForce RTX 309024 GBQ2_K21 GB~130

How these numbers are produced

  • Model shape comes from the model's own config.json — 32 layers, 32 attention heads, 8 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 (2 of 8), so throughput is scored on ~12.9B active parameters, not the full 47B. Memory is the opposite: every expert must still be resident in VRAM, so the figures above use the full weight set.Derived from architecture; matches the vendor-published active count within 1%.
  • Rental prices are pulled hourly from provider APIs (last updated Tue, 21 Jul 2026 15:20:05 GMT). Spot/interruptible pricing can change or vanish without notice.

Other models