01.AI on NVIDIA Ampere

Can I run Yi-1.5 34B on an RTX 3090?

Not at Q4_K_M — it needs 22.2 GB against 21.8 GB available. Drop to Q4_K_S and it fits, at about 29.9 tokens per second.

Won't fit

22.2 GB of 21.8 GB · 102%
023 GB
Weights 19.5 GB
KV cache 1.9 GB
Runtime overhead 0.9 GB
Over the limit 0.5 GB

Short by 0.5 GB. You can run it with 58 of 60 layers on the RTX 3090 and the rest in system RAM, at roughly 19.6 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.

Generation19.6tok/s
Prompt processing433tok/s
Max context6Ktokens
KV per 1K tokens0GB

Every quantisation of Yi-1.5 34B on a RTX 3090

Highlighted row is the highest quality that still fits at 8K context.

QuantisationbpwWeights TotalVerdictMax ctxtok/s
FP16 / BF16 16.00 64.1 GB 66.8 GB Won't fit 0.83
INT8 / W8A8 8.50 33.8 GB 36.6 GB Won't fit 2.36
Q8_0 (GGUF) 8.50 33.8 GB 36.6 GB Won't fit 2.36
FP8 (E4M3) 8.00 32.0 GB 34.8 GB Won't fit 2.65
Q6_K 6.56 26.3 GB 29.0 GB Won't fit 4.41
Q5_K_M 5.67 22.7 GB 25.5 GB Won't fit 7.54
Q5_K_S 5.52 22.1 GB 24.9 GB Won't fit 8.31
Q4_K_M 4.85 19.5 GB 22.2 GB Won't fit 6K 19.6
Q4_K_S 4.58 18.4 GB 21.2 GB Fits, but tight 10K 29.9
Q4_0 4.55 18.3 GB 21.1 GB Fits, but tight 11K 30.1
AWQ 4-bit 4.25 18.3 GB 21.0 GB Fits, but tight 11K 30.2
GPTQ 4-bit 4.25 18.3 GB 21.0 GB Fits, but tight 11K 30.2
MXFP4 4.25 18.3 GB 21.0 GB Fits, but tight 11K 30.2
IQ4_XS 4.25 17.2 GB 19.9 GB Fits, but tight 16K 32.0
Q3_K_M 3.91 15.8 GB 18.6 GB Runs comfortably 22K 34.5
IQ3_M 3.70 15.0 GB 17.8 GB Runs comfortably 25K 36.3
IQ3_XXS 3.06 12.5 GB 15.3 GB Runs comfortably 32K 42.9
Q2_K 2.63 10.8 GB 13.6 GB Runs comfortably 32K 48.9
IQ2_XXS 2.06 8.6 GB 11.4 GB Runs comfortably 32K 60.0
IQ1_M 1.75 7.4 GB 10.2 GB Runs comfortably 32K 68.5

Also worth checking