NVIDIA V100 32GB SXM2: what runs best?

RTX 4090 RTX 5090 RTX 3090 RTX 4060 Ti (16 GB)RTX 5070 Ti Radeon RX 7900 XTX M4 Max 128 GB M3 Ultra 512 GB DGX Spark (128 GB)All devices

NVIDIA · GPU

NVIDIA V100 32GB SXM2

Memory32 GB HBM2Bandwidth900 GB/s vendor specPlussystem RAM for offload (set below)

GPUs

V100 32GB SXM232 GB · 900 GB/s1

Custom GPU

System RAM

InstalledGBTypeSpeedMT/sChannels× 64-bit

≈ 89.6 GB/s theoretical ceiling (MT/s × 8 bytes × channels). DDR5’s two 32-bit subchannels count as one channel.

CPU

PlatformThreads for inference

Used for CPU layers/experts and the -t flag. A reference platform fills its maximum channels and speed; set what you actually populated.

Advanced: reserves, bandwidth, workload, models

Available VRAM per GPUGiB Runtime allowance per deviceGiB Spare margin per deviceGiB Unified memory for models% RAM kept for OS and appsGiB Measured RAM bandwidthGB/s GPU linkLanes Prompt length (first token)tokens Custom contexttokens Parallel sequences

Allow disk-streamed experts (documented cases)Include base checkpointsInclude specialistsInclude community & quant-trained variantsInclude models with unverified release dates

Defaults are editorial planning choices, not measurements: 1 GiB runtime + 0.5 GiB margin per device, 75% of unified memory, 8 GiB RAM for the OS, a vendor “24 GB” treated as 24 GiB (enter your driver’s available VRAM to override), 5 t/s minimum speed (research editorial default). Memory ledger

V100 32GB SXM264 GB DDR5-5600 ×2 ch · 56.0 GiB usable · 89.6 GB/s theoretical8K contextF16 KVany backend≥ 5 t/s

Best plans V100 32GB SXM2

Best quality at ≥ 5 t/s

Qwen3.8-27B

33.7 AA index Xhigh

Q8_0llama.cpp CUDAFull GPU

~18–26 t/s Estimated

Faster tradeoff

K2-Horizon-MoVA-36B-A4B

25.3 AA index single mode

Q5_K_Mllama.cpp CUDAFull GPU

~45–131 t/s Estimated extrapolated

Next point on the quality/speed front; then Qwen3.6-35B-A3B.

Possibly better · speed not established

No higher-scoring model has a feasible plan here.

WeightsKV cacheReserveFreeFits* = a state term unsizedMeasuredlinks to its sourceEstimatedcalibrated or extrapolated range, ~Ceilingupper bound, ≤Unknownnever zero

Meets your 5 t/s floor 36

Best plan per model that reaches your floor by a matching measurement, a calibrated range, or an extrapolated estimate (labelled). Sorted by the selected quality metric, then speed.

1
Qwen3.8-27B
Qwen/Qwen3.8-27B
2026-08-1427B dense
AA index33.7Xhigh
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits*
GPU
28.6 / 32.0 GiB
Generate~18–26 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
2
K2-Horizon-MoVA-36B-A4B
IFM/K2-Horizon-MoVA-36B-A4B
2026-09-0336B MoE · 4B active
AA index25.3single mode
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits
GPU
27.6 / 32.0 GiB
Generate~45–131 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
3
K2-Horizon-7B
IFM/K2-Horizon-7B
2026-09-039B tensors · dense
AA index20.6single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
11.5 / 32.0 GiB
Generate~50–73 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
4
Qwen3.6-35B-A3B
Qwen/Qwen3.6-35B-A3B
2026-04-1635B MoE · 3B active
AA index18.2Reasoning
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits*
GPU
25.8 / 32.0 GiB
Generate~74–213 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
5
Muse-Glimmer-30B
meta-models/Muse-Glimmer-30B
2026-08-1030B dense
AA index17.5High
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
29.3 / 32.0 GiB
Generate~17–25 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
6
K2-Horizon-3.7B
IFM/K2-Horizon-3.7B
2026-09-035.1B tensors · dense
AA index15.6single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
7.6 / 32.0 GiB
Generate~85–124 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
7
Qwen3.5-122B-A10B
Qwen/Qwen3.5-122B-A10B
2026-02-24122B MoE · 10B active
AA index15.6Reasoning
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU experts · 30 expert layers on CPUFits*
GPU
31.0 / 32.0 GiB
RAM
42.9 / 56.0 GiB
Generate~4.9–7.5 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
8
EXAONE-4.5-33B
LGAI-EXAONE/EXAONE-4.5-33B
2026-04-0933B dense
AA index13.2 est.Reasoning
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits
GPU
24.6 / 32.0 GiB
Generate~21–30 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
9
Qwen3.5-4B
Qwen/Qwen3.5-4B
2026-03-024B dense
AA index13.1 est.Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits*
GPU
6.1 / 32.0 GiB
Generate~108–156 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
10
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
2026-08-1130B MoE · 3B active
AA index12.9single mode
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits*
GPU
29.7 / 32.0 GiB
Generate~56–161 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
11
granite-4.2-30b
ibm-granite/granite-4.2-30b
2026-08-2530B dense
AA index12.7single mode
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits
GPU
22.8 / 32.0 GiB
Generate~23–34 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
12
MiniCPM5-2B
openbmb/MiniCPM5-2B
2026-09-072.5B dense
AA index12.5single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits
GPU
6.5 / 32.0 GiB
Generate~98–142 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
13
Hypernova-60B-2605
MultiverseComputingCAI/Hypernova-60B-2605
2026-05-0660B MoE · 4.8B active
Also: HyperNova-60B (listed with its dated update); Hypernova-60B-2602 (dated update, not separately evaluated)
AA index11.7 est.High
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU experts · 7 expert layers on CPUFits
GPU
31.8 / 32.0 GiB
RAM
7.8 / 56.0 GiB
Generate~13–23 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
14
Nemotron-Cascade-2-30B-A3B
nvidia/Nemotron-Cascade-2-30B-A3B
2026-03-1930B MoE · 3B active
AA index11.7 est.single mode
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits*
GPU
26.0 / 32.0 GiB
Generate~64–186 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
15
Qwen3.5-9B
Qwen/Qwen3.5-9B
2026-03-029B dense
AA index11.2Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits*
GPU
10.9 / 32.0 GiB
Generate~52–75 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
16
granite-4.2-8b
ibm-granite/granite-4.2-8b
2026-08-258B dense
AA index11.1single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
11.5 / 32.0 GiB
Generate~51–74 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
17
G9v3-3B
ai9stars/G9v3-3B
2026-07 (day unknown)3B dense
AA index10.8 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits
GPU
7.5 / 32.0 GiB
Generate~83–120 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
18
granite-4.2-3b
ibm-granite/granite-4.2-3b
2026-08-253B dense
AA index9.1single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
5.8 / 32.0 GiB
Generate~121–175 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
19
Tri-21B-Think
trillionlabs/Tri-21B-Think
2026-02 (day unknown)21B dense
AA index9.0 est.single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
23.6 / 32.0 GiB
Generate~22–32 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
20
LFM2.5-2.6B
LiquidAI/LFM2.5-2.6B
2026-08-042.7B dense
AA index8.4 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits*
GPU
6.7 / 32.0 GiB
Generate~94–135 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
21
LFM2.5-8B-A1B
LiquidAI/LFM2.5-8B-A1B
2026-05-288.3B MoE · 1.5B active
AA index7.2 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits*
GPU
17.4 / 32.0 GiB
Generate~55–158 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
22
Olmo-3.1-32B-ThinkSpecialist
allenai/Olmo-3.1-32B-Think
2025-12-1232B tensors · dense
AA index7.1 est.single mode
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits
GPU
24.0 / 32.0 GiB
Generate~21–31 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
23
Qwen3.5-2B
Qwen/Qwen3.5-2B
2026-03-022B dense
AA index6.9 est.Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits*
GPU
3.5 / 32.0 GiB
Generate~240–348 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
24
Olmo-3.1-32B-Instruct
allenai/Olmo-3.1-32B-Instruct
2025-12-1232B tensors · dense
AA index6.5 est.single mode
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits
GPU
24.0 / 32.0 GiB
Generate~21–31 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
25
Qwen3.5-0.8B
Qwen/Qwen3.5-0.8B
2026-03-020.8B dense
AA index6.1 est.Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits*
GPU
2.4 / 32.0 GiB
Generate~579–838 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
26
granite-4.0-h-small
ibm-granite/granite-4.0-h-small
2025-10-0232B tensors · MoE
AA index6.0 est.single mode
Plan & memory
Q5_K_Mllama.cpp CUDAFull GPUFits*
GPU
22.9 / 32.0 GiB
Generate~27–78 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
27
Phi-4Older
microsoft/phi-4
2024-12-1215B tensors · dense
AA index5.9 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits
GPU
30.4 / 32.0 GiB
Generate~17–25 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
28
Olmo-3-7B-ThinkSpecialist
allenai/Olmo-3-7B-Think
2025-11-207.3B tensors · dense
AA index5.6 est.single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
11.2 / 32.0 GiB
Generate~52–75 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
29
Olmo-3-7B-Instruct
allenai/Olmo-3-7B-Instruct
2025-11-207.3B tensors · dense
AA index5.2 est.single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
11.2 / 32.0 GiB
Generate~52–75 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
30
LFM2.5-1.2B-ThinkingSpecialist
LiquidAI/LFM2.5-1.2B-Thinking
2026-01-201.2B tensors · dense
AA index5.2 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits*
GPU
3.8 / 32.0 GiB
Generate~214–310 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
31
LFM2.5-1.2B-Instruct
LiquidAI/LFM2.5-1.2B-Instruct
2026-01-051.2B tensors · dense
AA index5.2 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits*
GPU
3.8 / 32.0 GiB
Generate~214–310 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
32
Molmo2-8BSpecialist
allenai/Molmo2-8B
2025-12-118.7B tensors · dense
AA index5.0 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits
GPU
17.9 / 32.0 GiB
Generate~30–44 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
33
LFM2.5-VL-1.6BSpecialist
LiquidAI/LFM2.5-VL-1.6B
2026-01-051.6B tensors · dense
AA index4.8 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits*
GPU
3.8 / 32.0 GiB
Generate~214–310 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
34
Apertus-8B-Instruct-2509
swiss-ai/Apertus-8B-Instruct-2509
2025-09-028.1B tensors · dense
AA index4.8 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUFits
GPU
17.5 / 32.0 GiB
Generate~31–45 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
35
NVIDIA-Nemotron-3-Nano-4B-BF16
nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16
2026-03-164B dense
AA index4.8single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits*
GPU
5.6 / 32.0 GiB
Generate~119–172 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
36
K2-Horizon-0.9B
IFM/K2-Horizon-0.9B
2026-09-031.1B tensors · dense
AA index3.0single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUFits
GPU
3.0 / 32.0 GiB
Generate~370–536 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown

+ 58 models without a AA index score

LFM2.5-230MFP16 · full gpu · ~1,213 t/s
LFM2.5-350MFP16 · full gpu · ~815 t/s
LittleLambFP16 · full gpu · ~585 t/s
LittleLamb-ToolCallingFP16 · full gpu · ~585 t/s
Jais-3-0.9BQ8_0 · full gpu · ~449 t/s
Qwen3-0.6BFP16 · full gpu · ~314 t/s
LFM2.5-1.2B-JPFP16 · full gpu · ~260 t/s
granite-4.0-h-tinyQ8_0 · full gpu · ~217 t/s
MiniCPM-V-4.7-35B-A3BQ5_K_M · full gpu · ~157 t/s
granite-4.1-3bQ8_0 · full gpu · ~157 t/s
LFM2.5-24B-A2BQ8_0 · full gpu · ~157 t/s
Qwen3.5-35B-A3BQ5_K_M · full gpu · ~150 t/s
Qwen3-1.7BFP16 · full gpu · ~137 t/s
Agents-A1-4BQ8_0 · full gpu · ~135 t/s
FrogNano-4B-2609Q8_0 · full gpu · ~131 t/s
Qwen3-4BQ8_0 · full gpu · ~127 t/s
Qwen3-30B-A3BQ5_K_M · full gpu · ~124 t/s
Qwen3-Coder-30B-A3B-InstructQ5_K_M · full gpu · ~124 t/s
Mellum2.1-12B-A2.5B-ThinkingQ8_0 · full gpu · ~124 t/s
AgentCPM-ReportQ4_K_M · full gpu · ~122 t/s
Jais-3-3.7BQ8_0 · full gpu · ~104 t/s
PaCoRe-8BQ8_0 · full gpu · ~67 t/s
Qwen3-8BQ8_0 · full gpu · ~67 t/s
AgentCPM-ExploreFP16 · full gpu · ~66 t/s
MiMo-V2.6-Distill-Qwen-9BQ8_0 · full gpu · ~64 t/s
Ornith-1.5-9BQ8_0 · full gpu · ~64 t/s
granite-4.1-8bQ8_0 · full gpu · ~62 t/s
Jais-3-7BQ8_0 · full gpu · ~61 t/s
Mixtral 8x7B Instruct v0.1Q4_K_M · full gpu · ~44 t/s
Mistral 7B Instruct v0.3FP16 · full gpu · ~41 t/s
IQuest-Coder-V1-7B-Base-Stage1FP16 · full gpu · ~40 t/s
IQuest-Coder-V1-7B-BaseFP16 · full gpu · ~40 t/s
IQuest-Coder-V1-7B-InstructFP16 · full gpu · ~40 t/s
IQuest-Coder-V1-7B-ThinkingFP16 · full gpu · ~40 t/s
MagenticBrainQ8_0 · full gpu · ~38 t/s
DeepSeek-R1-Distill-Llama-8BFP16 · full gpu · ~37 t/s
RLVR-8B-0926FP16 · full gpu · ~37 t/s
Nemotron-Cascade-8BFP16 · full gpu · ~37 t/s
NVIDIA-Nemotron-Nano-9B-v2-JapaneseFP16 · full gpu · ~35 t/s
Kolibri-1Q4_K_M · gpu + cpu experts · ~32 t/s
granite-4.1-30bQ5_K_M · full gpu · ~29 t/s
Olmo-3-32B-ThinkQ5_K_M · full gpu · ~26 t/s
Qwen3-32BQ5_K_M · full gpu · ~26 t/s
FrogBoss-32B-2510Q5_K_M · full gpu · ~26 t/s
Qwen3-Nemotron-32B-RLBFFQ5_K_M · full gpu · ~26 t/s
Mistral NeMo 12B Instruct (2407)FP16 · full gpu · ~25 t/s
Mistral Small 3.2 24B Instruct (2506)Q8_0 · full gpu · ~24 t/s
Jais-3-32BQ5_K_M · full gpu · ~24 t/s
Qwen3.6-27BQ8_0 · full gpu · ~22 t/s
IQuest-Coder-V1-14B-Base-Stage1FP16 · full gpu · ~21 t/s
IQuest-Coder-V1-14B-BaseFP16 · full gpu · ~21 t/s
IQuest-Coder-V1-14B-InstructFP16 · full gpu · ~21 t/s
IQuest-Coder-V1-14B-ThinkingFP16 · full gpu · ~21 t/s
Qwen3.5-27BQ8_0 · full gpu · ~21 t/s
IQuest-Coder-V1-40B-ThinkingQ5_K_M · full gpu · ~21 t/s
Qwen3-14BFP16 · full gpu · ~21 t/s
FrogMini-14B-2510FP16 · full gpu · ~21 t/s
Laguna-S-2.1Q5_K_M · gpu + cpu layers · ~5.4 t/s

Speed not established 7

Feasible plans with only a bandwidth ceiling (backends or placements without calibration). Check the evidence before relying on them.

1
G9v3-39A5B (preview)
ai9stars/G9v3-39A5B
2026-08 (day unknown)39B MoE · 5B active
AA index21.8 est.single mode
Plan & memory
native F32/BF16vLLMGPU streams weights from RAM · 43 GiB in RAMFits
GPU
32.0 / 32.0 GiB
RAM
42.4 / 56.0 GiB
Generate≤5.3 t/sCeilingnot established
First token · 4K prompt—Unknown
2
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16Specialist
nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16
2026-04-2830B MoE · 3B active
AA index10.3 est.single mode
Plan & memory
native F32/BF16vLLMGPU streams weights from RAM · 32 GiB in RAMFits*
GPU
32.0 / 32.0 GiB
RAM
31.0 / 56.0 GiB
Generate≤9.1 t/sCeilingnot established
First token · 4K prompt—Unknown
3
North-Mini-Code-1.0Specialist
CohereLabs/North-Mini-Code-1.0
2026-06-0930B MoE · 3B active
AA index9.9single mode
Plan & memory
native BF16vLLMGPU streams weights from RAM · 27 GiB in RAMFits
GPU
32.0 / 32.0 GiB
RAM
26.8 / 56.0 GiB
Generate≤10 t/sCeilingnot established
First token · 4K prompt—Unknown
4
MiniCPM5-1B
openbmb/MiniCPM5-1B
2026-05-191.1B dense
AA index8.8 est.Reasoning
Plan & memory
native BF16vLLMFull GPUFits
GPU
3.7 / 32.0 GiB
Generate≤398 t/sCeilingnot established
First token · 4K prompt—Unknown
5
Nanbeige4.1-3B
Nanbeige/Nanbeige4.1-3B
2026-02 (day unknown)3B dense
AA index8.4 est.single mode
Plan & memory
native BF16vLLMFull GPUFits
GPU
9.3 / 32.0 GiB
Generate≤111 t/sCeilingnot established
First token · 4K prompt—Unknown
6
Step3-VL-10B
stepfun-ai/Step3-VL-10B
2026-01 (day unknown)10B dense
AA index7.7 est.single mode
Plan & memory
native BF16vLLMFull GPUFits
GPU
21.6 / 32.0 GiB
Generate≤43 t/sCeilingnot established
First token · 4K prompt—Unknown
7
NVIDIA-Nemotron-Nano-12B-v2-VL-BF16Specialist
nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16
2025-10-2813B tensors · dense
AA index7.5 est.Reasoning
Plan & memory
native F32/BF16vLLMFull GPUFits*
GPU
26.2 / 32.0 GiB
Generate≤34 t/sCeilingnot established
First token · 4K prompt—Unknown

+ 4 models without a AA index score

Agents-A1native BF16 · gpu streams weights from ram · ≤ 7.3 t/s
MiniCPM-SALAnative BF16 · full gpu · ≤ 47 t/s
Penguin-VL-2Bnative BF16 · full gpu · ≤ 187 t/s
Penguin-VL-8Bnative BF16 · full gpu · ≤ 50 t/s

Memory not fully resolved 17

The weights fit, but the KV cache for this attention type (or the chosen cache layout) is not established, so a fit cannot be guaranteed.

1
Ling-3.0-flash-VL
inclusionAI/Ling-3.0-flash-VL
2026-09-09124B MoE · 5.5B active
AA index24.6single mode
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU experts · 24 expert layers on CPUKV unknown
GPU
31.9+ / 32.0 GiB
RAM
42.8 / 56.0 GiB
Generate~9.7–15 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
2
Ling-3.0-flash-Fin
inclusionAI/Ling-3.0-flash-Fin
2026 (month unknown)124B MoE · 5.1B active
AA index22.6single mode
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU experts · 24 expert layers on CPUKV unknown
GPU
31.6+ / 32.0 GiB
RAM
42.4 / 56.0 GiB
Generate~9.8–15 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
3
gemma-4-26B-A4B-it
google/gemma-4-26B-A4B-it
2026-04-0225B MoE · 3.8B active
AA index16.7 est.Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUKV unknown
GPU
26.5+ / 32.0 GiB
Generate~42–121 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
4
gemma-4-31B-it
google/gemma-4-31B-it
2026-04-0231B dense
AA index14.7Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUKV unknown
GPU
31.9+ / 32.0 GiB
Generate~16–23 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
5
gemma-4-12B-it
google/gemma-4-12B-it
2026-06-0312B dense
AA index14.2 est.Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUKV unknown
GPU
13.3+ / 32.0 GiB
Generate~40–58 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
6
LongCat-Flash-Lite
meituan-longcat/LongCat-Flash-Lite
2026-01 (day unknown)69B MoE · 3B active
AA index11.5 est.single mode
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU layers · 11/14 layers on GPUKV unknown
GPU
30.9+ / 32.0 GiB
RAM
8.0 / 56.0 GiB
Generate~37–54 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
7
Mistral-Small-4-119B-2603
mistralai/Mistral-Small-4-119B-2603
2026-03-16119B MoE · 6B active
AA index11.3Reasoning
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU experts · 21 expert layers on CPUKV unknown
GPU
30.7+ / 32.0 GiB
RAM
38.5 / 56.0 GiB
Generate~5.7–8.5 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
8
Ling-3.0-tiny
inclusionAI/Ling-3.0-tiny
2026-08 (day unknown)7.9B MoE · 1.3B active
AA index11.1single mode
Plan & memory
FP16llama.cpp CUDAFull GPUKV unknown
GPU
16.2+ / 32.0 GiB
Generate~65–189 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
9
Qwen3-Next-80B-A3B-InstructOlder
Qwen/Qwen3-Next-80B-A3B-Instruct
2025-09-1181B tensors · MoE
AA index9.6 est.single mode
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU experts · 17 expert layers on CPUKV unknown
GPU
31.5+ / 32.0 GiB
RAM
15.2 / 56.0 GiB
Generate~18–30 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
10
diffusiongemma-26B-A4B-it
google/diffusiongemma-26B-A4B-it
2026-06-1025B MoE · 3.8B active
AA index9.5 est.single mode
Plan & memory
Q8_0llama.cpp CUDAFull GPUKV unknown
GPU
26.5+ / 32.0 GiB
Generate~42–121 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
11
Qwen3-Coder-NextSpecialist
Qwen/Qwen3-Coder-Next
2026-02-0280B MoE · 3B active
AA index9.2single mode
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU layers · 32/48 layers on GPUKV unknown
GPU
31.8+ / 32.0 GiB
RAM
15.1 / 56.0 GiB
Generate~27–36 t/sEstimatedextrapolated≈ meets 5 t/s
First token · 4K prompt—Unknown
12
gemma-4-E4B-it
google/gemma-4-E4B-it
2026-04-028B dense
AA index8.9Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUKV unknown
GPU
9.0+ / 32.0 GiB
Generate~64–92 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
13
Falcon-H1R-7B
tiiuae/Falcon-H1R-7B
2026-01-057B dense
AA index7.8 est.single mode
Plan & memory
native BF16vLLMFull GPUKV unknown
GPU
15.6+ / 32.0 GiB
Generate≤59 t/sCeilingnot established
First token · 4K prompt—Unknown
14
gemma-4-E2B-it
google/gemma-4-E2B-it
2026-04-025.1B dense
AA index7.8 est.Reasoning
Plan & memory
Q8_0llama.cpp CUDAFull GPUKV unknown
GPU
6.1+ / 32.0 GiB
Generate~103–149 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
15
Llama 3.3 70B InstructOlderGated
meta-llama/Llama-3.3-70B-Instruct
2024-12-0671B tensors
AA index7.7 est.single mode
Plan & memory
AWQvLLMGPU streams weights from RAM · 7 GiB in RAMKV unknown
GPU
32.0+ / 32.0 GiB
RAM
6.5 / 56.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown
16
Phi-4-mini-instructOlder
microsoft/Phi-4-mini-instruct
2025-02-263.8B tensors · dense
AA index6.3 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUKV unknown
GPU
8.7+ / 32.0 GiB
Generate~67–96 t/sEstimatedmeets 5 t/s
First token · 4K prompt—Unknown
17
tiny-aya-globalGated
CohereLabs/tiny-aya-global
2026-02-173.3B tensors
AA index4.8 est.single mode
Plan & memory
FP16llama.cpp CUDAFull GPUKV unknown
GPU
7.7+ / 32.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown

+ 27 models without a AA index score

DeepSeek-R1-Distill-Qwen-1.5BQ8_0 · full gpu · ~328 t/s
NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16Q5_K_M · full gpu · ~169 t/s
GLM-4.7-FlashQ8_0 · full gpu · ~109 t/s
DeepSeek-R1-Distill-Qwen-7BFP16 · full gpu · ~41 t/s
Qwen2.5-7B-InstructFP16 · full gpu · ~41 t/s
DeepSeek-R1-Distill-Qwen-32BQ5_K_M · full gpu · ~27 t/s
Qwen2.5-32B-InstructQ5_K_M · full gpu · ~27 t/s
DeepSeek-R1-Distill-Qwen-14BFP16 · full gpu · ~21 t/s
Qwen2.5-14B-InstructFP16 · full gpu · ~21 t/s
Qwen2.5-Coder-7B-InstructFP16 · full gpu · ~20 t/s
Ling-3.0-flashQ4_K_M · gpu + cpu experts · ~13 t/s
DeepSeek-V3AWQ · full gpu · ≤ 12,707 t/s
Gemma 3 12B ITFP16 · full gpu
Gemma 3 1B ITnative BF16 · full gpu
Gemma 3 27B ITQ8_0 · full gpu
Gemma 3 4B ITFP16 · full gpu
Llama 2 7Bnative F32/F16 · full gpu
Llama 3.1 70B InstructQ5_K_M · cpu only
Llama 3.1 8B Instructnative BF16 · full gpu
Llama 3.2 1B InstructFP16 · full gpu
Llama 3.2 3B InstructFP16 · full gpu
Nemotron-Elastic-12Bnative BF16 · full gpu
Qwen2.5-Coder-32B-InstructAWQ · full gpu · ≤ 47 t/s
tiny-aya-earthFP16 · full gpu
tiny-aya-fireFP16 · full gpu
tiny-aya-waterFP16 · full gpu
Youtu-Parsing-Omninative BF16 · full gpu · ≤ 84 t/s

Speed unknown 7

No measurement, calibration or bandwidth figure for this plan.

1
gpt-oss-120bOlder
openai/gpt-oss-120b
2025-08-05117B tensors · MoE
AA index11.6High
Plan & memory
Q8_0llama.cpp CUDAGPU + CPU layers · 18/36 layers on GPUFits
GPU
31.2 / 32.0 GiB
RAM
29.7 / 56.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown
2
gpt-oss-20bOlder
openai/gpt-oss-20b
2025-08-0521B tensors · MoE
AA index9.0High
Plan & memory
native mxfp4vLLMFull GPUFits
GPU
14.5 / 32.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown
3
NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
2025-12-1532B tensors · MoE
AA index8.9Reasoning
Plan & memory
native F32/BF16vLLMGPU streams weights from RAM · 29 GiB in RAMFits*
GPU
32.0 / 32.0 GiB
RAM
28.4 / 56.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown
4
granite-4.0-h-1b
ibm-granite/granite-4.0-h-1b
2025-10-281.5B tensors · MoE
AA index5.2 est.single mode
Plan & memory
native BF16vLLMFull GPUFits*
GPU
4.3 / 32.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown
5
granite-4.0-micro
ibm-granite/granite-4.0-micro
2025-10-023.4B tensors · MoE
AA index5.1 est.single mode
Plan & memory
native BF16vLLMFull GPUFits
GPU
8.5 / 32.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown
6
granite-4.0-350m
ibm-granite/granite-4.0-350m
2025-10-280.4B tensors · MoE
AA index4.8 est.single mode
Plan & memory
native BF16vLLMFull GPUFits
GPU
2.4 / 32.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown
7
granite-4.0-h-350m
ibm-granite/granite-4.0-h-350m
2025-10-280.3B tensors · MoE
AA index4.8 est.single mode
Plan & memory
native BF16vLLMFull GPUFits*
GPU
2.2 / 32.0 GiB
Generate—Unknownspeed unknown
First token · 4K prompt—Unknown

+ 5 models without a AA index score

granite-4.0-1bnative BF16 · full gpu
granite-4.0-h-micronative BF16 · full gpu
Jais-3-MoVA-36B-A4BQ5_K_M · full gpu
Nemotron-3-Super-120B-A12B-BF16-MTPv2native F32/BF16 · full gpu
NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-BF16Q5_K_M · full gpu

Below 5 t/s 2

Fits, but the best evidence (measured, estimated or the bandwidth ceiling) is under your floor. Lower the floor or add memory bandwidth.

1
Mistral-Medium-3.5-128B
mistralai/Mistral-Medium-3.5-128B
2026-04-28128B dense
AA index14.2single mode
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU layers · 37/88 layers on GPUFits
GPU
32.0 / 32.0 GiB
RAM
42.0 / 56.0 GiB
Generate~0.4–0.5 t/sEstimatedextrapolatedbelow 5 t/s
First token · 4K prompt—Unknown
2
Apertus-70B-Instruct-2509
swiss-ai/Apertus-70B-Instruct-2509
2025-09-0271B tensors · dense
AA index5.1 est.single mode
Plan & memory
Q4_K_Mllama.cpp CUDAGPU + CPU layers · 56/80 layers on GPUFits
GPU
31.8 / 32.0 GiB
RAM
13.0 / 56.0 GiB
Generate~1.4–1.7 t/sEstimatedextrapolatedbelow 5 t/s
First token · 4K prompt—Unknown

+ 7 models without a AA index score

GLM-4.5-AirQ4_K_M · gpu + cpu experts · ~3.4 t/s
Qwen2.5-72B-InstructQ4_K_M · gpu + cpu layers · ~1.8 t/s
DeepSeek-R1-Distill-Llama-70BQ4_K_M · gpu + cpu layers · ~1.6 t/s
gpt-oss-puzzle-88Bnative mxfp4 · gpu streams weights from ram · ≤ 1.7 t/s
IQuest-Coder-V1-40B-Loop-Thinkingnative BF16 · gpu streams weights from ram · ≤ 0.6 t/s
K2-Horizon-32Bnative BF16 · gpu streams weights from ram · ≤ 0.8 t/s
Mixtral 8x22B Instruct v0.1AWQ · gpu streams weights from ram · ≤ 2.6 t/s

67 models with no feasible plan on this setup

MiMo-V2.6-Pro-MOPDnative fp8 527 GiB
GLM-5.3native fp8 704 GiB
Step-5-Previewnative fp8 600 GiB
Kimi-K3native F32/BF16 1,454 GiB
GLM-5.3-Flashnative fp8 306 GiB
Ling-3.1-flashQ4_K_M 333 GiB
Qwen3.8-2.4T-A95Bnative BF16 4,556 GiB
Qwen3.8-Flash-NextQ4_K_M 111 GiB
DeepSeek-V4.1-FlashQ4_K_M 414 GiB
Muse-Spark-Open-816B-A42BQ4_K_M 465 GiB exceeds memory
MiMo-V2.6-Flash-MOPDnative fp8 161 GiB
DeepSeek-V4-Pro-0813native fp8 831 GiB
Motif-3native BF16 586 GiB
K2-Horizon-375B-A23Bnative BF16 706 GiB
Beam-1.0-501B-A23BQ5_K_M 331 GiB
MiniMax-M3Q4_K_M 247 GiB
Nex-N2-ProQ4_K_M 225 GiB
NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16native BF16/F32 232 GiB
Kimi-K2.7-Codenative BF16/F32 554 GiB
Inkling-SmallQ4_K_M 151 GiB
Hy3Q4_K_M 170 GiB
InklingQ8_0 798 GiB
Solar-Open2-250Bnative BF16/F32 466 GiB
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16native BF16/F32 1,044 GiB
A.X-K2native fp8 647 GiB
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16native F32/BF16 139 GiB
K-EXAONE-2.0-750B-A37Bnative F32/BF16 1,396 GiB
LongCat-2.0native BF16/F32 3,308 GiB
Qwen3.5-397B-A17BGPTQ 220 GiB
Ring-2.6-1Tnative quantized 970 GiB
command-a-plus-05-2026-bf16native BF16 407 GiB
Qwen3-Next-80B-A3B-Thinkingnative BF16 151 GiB
Trinity-Large-ThinkingQ4_K_M 225 GiB
Llama 4 Maverick 17B-128E InstructQ4_K_M 226 GiB exceeds memory
sarvam-105bnative F32 395 GiB
Llama 4 Scout 17B-16E InstructQ4_K_M 62.9 GiB exceeds memory
Agnes-2.5-Pro-Alphanative BF16 739 GiB
Agnes-3.0-Qwennative BF16/I64 335 GiB
Apertus-v1.1-0.5B-InstructContext 8K exceeds this model's configured maximum (4K)
Apertus-v1.1-1.5B-InstructContext 8K exceeds this model's configured maximum (4K)
Apertus-v1.1-4B-InstructContext 8K exceeds this model's configured maximum (4K)
Apertus-v1.5-8Bno backend/format combination
BaguettotronContext 8K exceeds this model's configured maximum (4K)
DeepSeek-R1Q4_K_M 377 GiB
DeepSeek-V4-Flash-0731native fp8 155 GiB
GLM-5.2Q4_K_M 434 GiB
Hy3-previewnative BF16/F32 557 GiB
Hy4-previewnative BF16/F32 1,453 GiB
Jais-3-375B-A23Bnative BF16/F32 706 GiB
Kimi-K2.6native F32/I32 554 GiB
Ming-flash-omni-2.0native BF16 194 GiB
Ming-flash-omni-Previewnative BF16/U8 194 GiB
MiniMax-M2.7Q4_K_M 131 GiB
MonadContext 8K exceeds this model's configured maximum (2K)
NVIDIA-Nemotron-3-Super-120B-A12B-BF16native F32/BF16 230 GiB
NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4native modelopt 328 GiB
NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16native BF16/F32 146 GiB
NVIDIA-Nemotron-Labs-Teacher-Chatnative F32/BF16 1,044 GiB
NVIDIA-Nemotron-Labs-Teacher-Competition-Codingnative BF16/F32 1,044 GiB
NVIDIA-Nemotron-Labs-Teacher-Instruction-Followingnative BF16/F32 1,044 GiB
NVIDIA-Nemotron-Labs-Teacher-STEMnative BF16/F32 1,044 GiB
Qwen3-235B-A22BGPTQ 116 GiB
Qwen3-Coder-480B-A35B-InstructQ4_K_M 270 GiB
Qwen3-VL-235B-A22B-InstructQ4_K_M 132 GiB
Qwen3-VL-235B-A22B-ThinkingQ4_K_M 132 GiB
Step-3.7-FlashQ4_K_M 114 GiB
Apertus-v1.5-70Bno verified artifact for these backends

Specs & sources

Memory	32 GB HBM2	source
Bandwidth	900 GB/s	source
FP16	unknown	No unambiguous comparable value verified; never substitute AI TOPS, BF16, or whole-system power.
Power	300 W	source

Vendor calls throughput Deep Learning TFLOPS without precision in this page; FP16 field left unknown.

Reported speeds

0 observations

No reported observation resolved to this exact device yet.

Open in speed database Submit yours

Doesn't fit? Rent a GPU

RunpodPlaceholderCloud GPU pods you can rent for running larger models.
Vast.aiPlaceholderGPU rental marketplace with many hosts and GPU types.
LambdaPlaceholderCloud GPU instances for AI training and inference.
TensorDockPlaceholderGPU cloud marketplace with consumer and datacenter GPUs.

Sponsored placeholder links: LocalFit may earn a commission if these become affiliate links. They never influence fit results, estimates or picks.