DATA-DRIVEN MODEL COMPARISON

Best local LLMs for 64 and 128 GB RAM or VRAM

127 catalog weight sets fit the 44.8 GB system-memory planning budget. “Best” means a trade-off between a published benchmark score and weight size, not a tested run on your computer.

Data snapshot: 14 October 2026. Scores and weight sizes retain their original sources.

Check your own memory budget

64 GB and 128 GB: weight budgets

Illustrative reserves for one local model. These are not measured requirements.
Memory pool	Capacity	Weight budget	Reserved space	Candidate files
System RAM / unified memory	64 GB	44.8 GB	19.2 GB	127
Dedicated VRAM	64 GB	51.2 GB	12.8 GB	128
System RAM / unified memory	128 GB	89.6 GB	38.4 GB	129
Dedicated VRAM	128 GB	102.4 GB	25.6 GB	129

The 64 GB budget covers 127 catalog files; the 128 GB budget covers 129. 2 additional files fit the 128 GB budget; see their sizes below. Moving to 128 GB leaves 64 GB more physical memory for weights, context, software or concurrent sessions. Actual usable GPU memory depends on the system and runner.

First Pareto choice on MMLU-Pro

Qwen3.5-4B · 79.1 / 100 · 2.741 GB · Q4_K_M. This choice is calculated within this page’s eligible set.

MMLU-Pro: models within the RAM budget

Knowledge and reasoning across many subjects. Published accuracy; different evaluation settings can change the result.

MMLU-Pro · GB = 1 billion bytes. Published scores may use different settings or precision.
Model / quantization	Weights	Parameters	Score / 100	Score / GB	Budget left	Sources
Qwen3.5-4BQ4_K_M · Pareto choice	2.741 GB	4.206B	79.1	28.859	42.059 GB	Get weights Score source
MiniCPM5 2BQ4_K_M · Pareto choice	1.561 GB	2.517B	70.8	45.346	43.239 GB	Get weights Score source
Qwen3.5-9BQ4_K_M · Pareto choice	5.681 GB	9.653B	82.5	14.523	39.119 GB	Get weights Score source
LFM2.5 230MQ4_K_M · Pareto choice	0.153 GB	0.23B	20.25	132.002	44.647 GB	Get weights Score source
Qwen3.5-27BQ4_K_M · Pareto choice	16.741 GB	27.781B	86.1	5.143	28.059 GB	Get weights Score source
Qwen3.6-27BQ4_K_M · Pareto choice	16.817 GB	27.781B	86.2	5.126	27.983 GB	Get weights Score source
Qwen3.5-35B-A3BQ4_K_M	22.285 GB	35.952B	85.3	3.828	22.515 GB	Get weights Score source
gemma-4-31B-itQ4_K_M	18.324 GB	31.273B	85.2	4.65	26.476 GB	Get weights Score source
Qwen3.6-35B-A3BQ4_K_M	22.135 GB	35.952B	85.2	3.849	22.665 GB	Get weights Score source
Swift-Qwen3.8-27bQ4_K_M	17.442 GB	27.781B	84.95	4.87	27.358 GB	Get weights Score source
Seed-OSS-36B-InstructQ4_K_M	21.762 GB	36.151B	82.7	3.8	23.038 GB	Get weights Score source
gemma-4-26B-A4B-itQ4_K_M	16.948 GB	25.806B	82.6	4.874	27.852 GB	Get weights Score source
NVIDIA-Nemotron-3.5-Lightning-30B-A3BQ4_K_M	25.477 GB	32.913B	81.94	3.216	19.323 GB	Get weights Score source
Qwen3-30B-A3B-Thinking-2507Q4_K_M	18.557 GB	30.532B	80.9	4.36	26.243 GB	Get weights Score source
Qwen3-30B-A3B-Instruct-2507Q4_K_M	18.557 GB	30.532B	78.4	4.225	26.243 GB	Get weights Score source
NVIDIA-Nemotron-3-Nano-30B-A3BQ4_K_M	24.574 GB	31.578B	78.3	3.186	20.226 GB	Get weights Score source
gemma-4-12B-itQ4_K_M	7.122 GB	11.96B	77.2	10.84	37.678 GB	Get weights Score source
Granite 4.2 8BQ4_K_M	5.348 GB	8.792B	74.04	13.845	39.452 GB	Get weights Score source
Tiel-Coder 35B-A3BUD-Q4_K_XL	22.36 GB	34.661B	73.7	3.296	22.44 GB	Get weights Score source
Gemma 4 E4BQ4_K_M	4.977 GB	7.996B	69.4	13.944	39.823 GB	Get weights Score source
Granite 4.2 3BQ4_K_M	2.244 GB	3.66B	67.84	30.232	42.556 GB	Get weights Score source
LFM2.5 2.6BQ4_K_M	1.674 GB	2.697B	65.2	38.938	43.126 GB	Get weights Score source
Gemma 4 E2BQ4_K_M	3.107 GB	5.123B	60	19.313	41.693 GB	Get weights Score source
Noema 2BQ8_0	2.012 GB	1.882B	51.1	25.397	42.788 GB	Get weights Score source
Kernel Logic 1.5Q4_K_M	4.683 GB	7.616B	38	8.114	40.117 GB	Get weights Score source
Midnight-Miqu-70B-v1.5Q4_K_M	41.423 GB	68.977B	31.39	0.758	3.377 GB	Get weights Score source

LiveCodeBench v6: models within the RAM budget

Solving coding problems. Version 6 only; published pass rates can vary with generation settings.

LiveCodeBench v6 · GB = 1 billion bytes. Published scores may use different settings or precision.
Model / quantization	Weights	Parameters	Score / 100	Score / GB	Budget left	Sources
Nanbeige 4.2 3BQ4_K_M · Pareto choice	2.684 GB	4.17B	72.5	27.012	42.116 GB	Get weights Score source
MiniCPM5 2BQ4_K_M · Pareto choice	1.561 GB	2.517B	69.1	44.257	43.239 GB	Get weights Score source
Granite 4.2 3BQ4_K_M · Pareto choice	2.244 GB	3.66B	69.71	31.065	42.556 GB	Get weights Score source
Granite 4.2 8BQ4_K_M · Pareto choice	5.348 GB	8.792B	73.24	13.695	39.452 GB	Get weights Score source
Mellum2.1 12B A2.5B ThinkingQ4_K_M · Pareto choice	8.071 GB	12.15B	82	10.159	36.729 GB	Get weights Score source
Qwen3.8-27B-GSQ-RCO-GGUFIQ3_XXS · Pareto choice	10.094 GB	26.896B	84.57	8.378	34.706 GB	Get weights Score source
K2 Horizon 0.9BQ4_K_M · Pareto choice	0.666 GB	1.078B	37.4	56.141	44.134 GB	Get weights Score source
Qwen3.8-27BQ4_K_M · Pareto choice	16.464 GB	27.781B	90.3	5.485	28.336 GB	Get weights Score source
Qwen3.6-27BQ4_K_M	16.817 GB	27.781B	83.9	4.989	27.983 GB	Get weights Score source
Swift 1.5 Qwen3.8-27BQ4_K_M	17.442 GB	27.321B	81.71	4.685	27.358 GB	Get weights Score source
Swift-Qwen3.8-27bQ4_K_M	17.442 GB	27.781B	81.55	4.675	27.358 GB	Get weights Score source
Qwen3.5-27BQ4_K_M	16.741 GB	27.781B	80.7	4.821	28.059 GB	Get weights Score source
Qwen3.6-35B-A3BQ4_K_M	22.135 GB	35.952B	80.4	3.632	22.665 GB	Get weights Score source
gemma-4-31B-itQ4_K_M	18.324 GB	31.273B	80	4.366	26.476 GB	Get weights Score source
gemma-4-26B-A4B-itQ4_K_M	16.948 GB	25.806B	77.1	4.549	27.852 GB	Get weights Score source
Qwen3.5-35B-A3BQ4_K_M	22.285 GB	35.952B	74.6	3.348	22.515 GB	Get weights Score source
gemma-4-12B-itQ4_K_M	7.122 GB	11.96B	72	10.11	37.678 GB	Get weights Score source
Seed-OSS-36B-InstructQ4_K_M	21.762 GB	36.151B	67.4	3.097	23.038 GB	Get weights Score source
Qwen3-30B-A3B-Thinking-2507Q4_K_M	18.557 GB	30.532B	66	3.557	26.243 GB	Get weights Score source
Qwen3.5-9BQ4_K_M	5.681 GB	9.653B	65.6	11.548	39.119 GB	Get weights Score source
GLM-4.7-FlashQ4_K_M	18.312 GB	31.221B	64	3.495	26.488 GB	Get weights Score source
Agents A1 4BQ4_K_M	2.709 GB	4.539B	59.6	22.002	42.091 GB	Get weights Score source
NeoHorse 1 4BQ4_K_M	2.709 GB	4.206B	59.43	21.94	42.091 GB	Get weights Score source
LFM2.5 2.6BQ4_K_M	1.674 GB	2.697B	59.41	35.48	43.126 GB	Get weights Score source
Qwen3.5-4BQ4_K_M	2.741 GB	4.206B	55.8	20.358	42.059 GB	Get weights Score source
Gemma 4 E4BQ4_K_M	4.977 GB	7.996B	52	10.448	39.823 GB	Get weights Score source
Gemma 4 E2BQ4_K_M	3.107 GB	5.123B	44	14.163	41.693 GB	Get weights Score source
Qwen3-30B-A3B-Instruct-2507Q4_K_M	18.557 GB	30.532B	43.2	2.328	26.243 GB	Get weights Score source

Additional candidates for 128 GB

128 GB additional files · GB = 1 billion bytes. Published scores may use different settings or precision.
Model / quantization	Weights	Parameters	Score / 100	Score / GB	Budget left	Sources
Aleph Alpha Kolibri-1Q4_K_M · Pareto choice	47.454 GB	78.103B	80	1.686	42.146 GB	Get weights Score source
Xor 1.1BF16	70.214 GB	35.107B	Not recorded	—	19.386 GB	Get weights

All 129 model files within this tier’s RAM budget

Some files have scores only on other benchmarks. This complete size-ordered list keeps those candidates visible without assigning missing scores a zero. The table covers the 128 GB budget; use the 64 GB tables above for the smaller pool.

Show every candidate, including models without MMLU-Pro

All memory candidates · MMLU-Pro where recorded · GB = 1 billion bytes. Published scores may use different settings or precision.
Model / quantization	Weights	Parameters	Score / 100	Score / GB	Budget left	Sources
Vela Lumen v1.1F16	0.143 GB	0.048B	Not recorded	—	89.457 GB	Get weights
LFM2.5 230MQ4_K_M · Pareto choice	0.153 GB	0.23B	20.25	132.002	89.447 GB	Get weights Score source
PicoLM 80MF16	0.18 GB	0.09B	Not recorded	—	89.42 GB	Get weights
PicoLM v2 81MF16	0.193 GB	0.096B	Not recorded	—	89.407 GB	Get weights
PicoLM v2.1 81MF16	0.193 GB	0.096B	Not recorded	—	89.407 GB	Get weights
Udgam 384M ChatQ4_K_M	0.238 GB	0.384B	Not recorded	—	89.362 GB	Get weights
Quill 1 PreviewQ4_K_MCode-completion model; not a general chat model.	0.398 GB	0.494B	Not recorded	—	89.202 GB	Get weights
Liquid AI d1-omni-600MQ8_0	0.407 GB	0.587B	Not recorded	—	89.193 GB	Get weights
LightOnOCR-3 0.8BQ4_K_M	0.529 GB	0.853B	Not recorded	—	89.071 GB	Get weights
Satella-RAG 0.8BQ5_K_M	0.593 GB	0.773B	Not recorded	—	89.007 GB	Get weights
K2 Horizon 0.9BQ4_K_M	0.666 GB	1.078B	Not recorded	—	88.934 GB	Get weights
MiniCPM-V 4.7 1BQ4_K_M	0.672 GB	1.006B	Not recorded	—	88.928 GB	Get weights
OneJev 1.5 0.8BQ4_K_M	0.672 GB	1.007B	Not recorded	—	88.928 GB	Get weights
d3-edge 0.6BQ8_0	0.806 GB	0.748B	Not recorded	—	88.794 GB	Get weights
Switchboard Router 0.5B v0.1F16	0.994 GB	0.494B	Not recorded	—	88.606 GB	Get weights
Spark X2.5 1.7BQ4_K_M	1.107 GB	1.708B	Not recorded	—	88.493 GB	Get weights
Curious 2BQ4_0	1.215 GB	1.882B	Not recorded	—	88.385 GB	Get weights
Noema 1.5 2BQ4_K_M	1.274 GB	1.882B	Not recorded	—	88.326 GB	Get weights
StartLux-Decision-v1-2BQ4_K_M	1.274 GB	1.882B	Not recorded	—	88.326 GB	Get weights
Arm-optimized Qwen3.5-2BQ4_K_M	1.274 GB	1.882B	Not recorded	—	88.326 GB	Get weights
Empero Qwen3.8 2BQ4_K_M	1.312 GB	2.274B	Not recorded	—	88.288 GB	Get weights
Index Translate 2BQ4_K_MTranslation specialist, not a general chat ranking. Apache-2.0 licensed. The official llama.cpp recipe uses the chat template with thinking disabled, temperature 0 and the target-language prompt. This 1.312 GB Q4_K_M file is text-capable by itself; image input needs the separate compatible projector. Runtime, KV cache and long contexts need memory beyond the weight file. The recorded FLORES result is from the source checkpoint, not a Q4 speed or compatibility test.	1.312 GB	2.274B	Not recorded	—	88.288 GB	Get weights
MiniCPM5 abliterated v2Q4_K_M	1.561 GB	2.517B	Not recorded	—	88.039 GB	Get weights
MiniCPM5 2BQ4_K_M · Pareto choice	1.561 GB	2.517B	70.8	45.346	88.039 GB	Get weights Score source
ColGrep-Agent-2BQ4_K_M	1.561 GB	2.517B	Not recorded	—	88.039 GB	Get weights
LFM2.5 2.6BQ4_K_M	1.674 GB	2.697B	65.2	38.938	87.926 GB	Get weights Score source
DFM MimirQ8_0	1.913 GB	1.787B	Not recorded	—	87.687 GB	Get weights
TwIL LM3Q4_K_M	1.915 GB	3.075B	Not recorded	—	87.685 GB	Get weights
Noema 2BQ8_0	2.012 GB	1.882B	51.1	25.397	87.588 GB	Get weights Score source
Josev-2BQ8_0	2.045 GB	1.912B	Not recorded	—	87.555 GB	Get weights
Limite 1B - ViolettoBF16BF16 Safetensors; custom code or vLLM plugin. Single-turn math specialist.	2.071 GB	1B	Not recorded	—	87.529 GB	Get weights
Granite 4.2 3BQ4_K_M	2.244 GB	3.66B	67.84	30.232	87.356 GB	Get weights Score source
webAI TwIL-LM3-ProQ4_K_M	2.244 GB	3.66B	Not recorded	—	87.356 GB	Get weights
OpenDecider-small-tdQ4_K_M	2.497 GB	4.022B	Not recorded	—	87.103 GB	Get weights
Clary 2 Cyber HeavyQ4_K_M	2.497 GB	4.022B	Not recorded	—	87.103 GB	Get weights
RefinedToolCall v5Q6_K	2.538 GB	3.086B	Not recorded	—	87.062 GB	Get weights
grug 3BQ4_K_M	2.575 GB	4.17B	Not recorded	—	87.025 GB	Get weights
Spark X2.5 4BQ4_K_M	2.6 GB	4.112B	Not recorded	—	87 GB	Get weights
Granite concision editF16	2.673 GB	1.335B	Not recorded	—	86.927 GB	Get weights
Nanbeige 4.2 3BQ4_K_M	2.684 GB	4.17B	Not recorded	—	86.916 GB	Get weights
Wald-Q4B v2.1Q4_K_M	2.709 GB	4.66B	Not recorded	—	86.891 GB	Get weights
NeoHorse 1 4BQ4_K_M	2.709 GB	4.206B	Not recorded	—	86.891 GB	Get weights
APUS Reflex 4BQ4_K_M	2.709 GB	4.206B	Not recorded	—	86.891 GB	Get weights
Afrique Qwen v2Q4_K_M	2.709 GB	4.206B	Not recorded	—	86.891 GB	Get weights
Agents A1 4BQ4_K_M	2.709 GB	4.539B	Not recorded	—	86.891 GB	Get weights
Qwen3.5-4BQ4_K_M · Pareto choice	2.741 GB	4.206B	79.1	28.859	86.859 GB	Get weights Score source
Empero Qwen3.8 4BQ4_K_M	2.783 GB	4.66B	Not recorded	—	86.817 GB	Get weights
FrogNano-4B-2609Q4_K_M	2.798 GB	4.326B	Not recorded	—	86.802 GB	Get weights
Liquid AI d1-3BQ8_0Decision model, not a chat model. Use a llama.cpp build with /v1/systemone support. The selected Q8_0 file contains language weights; image input needs a compatible separate projector. Total parameters include the vision encoder. LFM Open License v1.0 has commercial-use conditions. The recorded score is not a Q8_0 compatibility or speed test.	2.875 GB	3.123B	Not recorded	—	86.725 GB	Get weights
Gemma 4 E2BQ4_K_M	3.107 GB	5.123B	60	19.313	86.493 GB	Get weights Score source
K2 Horizon 3.7BQ4_K_M	3.157 GB	5.058B	Not recorded	—	86.443 GB	Get weights
Bonsai 27B (1-bit)Q1_0Custom Prism llama.cpp kernels required; check backend support.	3.803 GB	26.896B	Not recorded	—	85.797 GB	Get weights
Switchboard Router 4BQ8_0	4.28 GB	4.022B	Not recorded	—	85.32 GB	Get weights
Quinto 0.1 4BQ8_0	4.482 GB	4.206B	Not recorded	—	85.118 GB	Get weights
H2O-Lightning 4B NativeQ8_0	4.482 GB	4.206B	Not recorded	—	85.118 GB	Get weights
Kernel Logic 1.5Q4_K_M	4.683 GB	7.616B	38	8.114	84.917 GB	Get weights Score source
Ling 3.0 tinyQ4_K_M	4.824 GB	7.893B	Not recorded	—	84.776 GB	Get weights
Gemma 4 E4BQ4_K_M	4.977 GB	7.996B	69.4	13.944	84.623 GB	Get weights Score source
Ai2 AstaBrief 8BQ4_K_M	5.028 GB	8.191B	Not recorded	—	84.572 GB	Get weights
DuoNeural Hermes v3Q4_K_M	5.156 GB	8.468B	Not recorded	—	84.444 GB	Get weights
DuoNeural Hermes v2Q4_K_M	5.156 GB	8.468B	Not recorded	—	84.444 GB	Get weights
UltraCoder L3Q4_K_M	5.156 GB	8.468B	Not recorded	—	84.444 GB	Get weights
Granite 4.2 8BQ4_K_M	5.348 GB	8.792B	74.04	13.845	84.252 GB	Get weights Score source
K2 Horizon 7BQ4_K_M	5.592 GB	8.999B	Not recorded	—	84.008 GB	Get weights
Qwythos-9B Claude Mythos 5 1MQ4_K_M	5.629 GB	8.954B	Not recorded	—	83.971 GB	Get weights
RapidChat 9BQ4_K_M	5.629 GB	9.41B	Not recorded	—	83.971 GB	Get weights
APUS Reflex 9BQ4_K_M	5.629 GB	9.41B	Not recorded	—	83.971 GB	Get weights
OpenMobile-2 9BQ4_K_M	5.629 GB	8.954B	Not recorded	—	83.971 GB	Get weights
Qwen3.5-9BQ4_K_M · Pareto choice	5.681 GB	9.653B	82.5	14.523	83.919 GB	Get weights Score source
Empero Qwen3.8 9BQ4_K_MReasoning distillation, not an official Qwen 3.8 9B release. Use a recent llama.cpp build with Qwen3.5 / Gated DeltaNet support; older builds can fail to load it. Apache-2.0 licensed. The 5.780 GB Q4_K_M file is only the weight file: runtime, KV cache and long contexts need additional memory. The recorded MMLU and GSM8K results are from the source checkpoint, not Q4 speed or compatibility tests.	5.78 GB	9.653B	Not recorded	—	83.82 GB	Get weights
Index-Translate-9BQ4_K_M	5.78 GB	9.653B	Not recorded	—	83.82 GB	Get weights
Ornith-1.5-9BQ4_K_M	5.78 GB	9.197B	Not recorded	—	83.82 GB	Get weights
MiMo-V2.6-Distill-Qwen-9BQ4_K_M	5.841 GB	8.954B	Not recorded	—	83.759 GB	Get weights
Ternary Bonsai 2 27BPTQ1_0Custom Prism llama.cpp kernels required; check backend support.	5.947 GB	26.896B	Not recorded	—	83.653 GB	Get weights
Ternary Bonsai 2 27B Abliterated (BoldingBuilds)PTQ1_0	5.947 GB	26.896B	Not recorded	—	83.653 GB	Get weights
Mitsuba-ComfyUI-27BPTQ1_0	5.998 GB	27.321B	Not recorded	—	83.602 GB	Get weights
Cloudflare Clef-flash 9BQ4_K_MDecision model. llama.cpp uses /v1/systemone; text generation is unavailable. Image input requires current model support plus a compatible multimodal projector. See the Clef comparison for sources and checks.	6.041 GB	9.076B	Not recorded	—	83.559 GB	Get weights
gemma-4-12B-itQ4_K_M	7.122 GB	11.96B	77.2	10.84	82.478 GB	Get weights Score source
Ternary Bonsai 27BPQ2_0Custom Prism llama.cpp kernels required; check backend support.	7.165 GB	26.896B	Not recorded	—	82.435 GB	Get weights
Swift Bonsai 2 27B (PQ2_0)PQ2_0	7.206 GB	26.896B	Not recorded	—	82.394 GB	Get weights
Mistral-Nemo-Instruct-2407Q4_K_M	7.477 GB	12.248B	Not recorded	—	82.123 GB	Get weights
Mellum2.1 12B A2.5B ThinkingQ4_K_M	8.071 GB	12.15B	Not recorded	—	81.529 GB	Get weights
Gmcoder 9BQ8_0	9.786 GB	8.954B	Not recorded	—	79.814 GB	Get weights
Qwen3.8-27B-GSQ-RCO-GGUFIQ3_XXS	10.094 GB	26.896B	Not recorded	—	79.506 GB	Get weights
Humanizer 12B RLRt2Q8_0	12.67 GB	11.907B	Not recorded	—	76.93 GB	Get weights
Winnow 12BQ8_0	12.67 GB	11.96B	Not recorded	—	76.93 GB	Get weights
GLM-4.7-Flash-REAP-23B-A3BQ4_K_M	14.114 GB	22.996B	Not recorded	—	75.486 GB	Get weights
LFM2.5-24B-A2BQ4_K_M	14.526 GB	23.972B	Not recorded	—	75.074 GB	Get weights
Qwen3.8-27BQ4_K_M	16.464 GB	27.781B	Not recorded	—	73.136 GB	Get weights
OpenMobile-2 27BQ4_K_M	16.547 GB	26.896B	Not recorded	—	73.053 GB	Get weights
OpenJev 27BQ4_K_M	16.547 GB	26.896B	Not recorded	—	73.053 GB	Get weights
ThakiCloud RAG-Gate-27BQ4_K_M	16.547 GB	27.321B	Not recorded	—	73.053 GB	Get weights
Lumma-Fev-27BQ4_K_M	16.547 GB	26.085B	Not recorded	—	73.053 GB	Get weights
Qwen3.8 27B Humanlike Chat 2.0Q4_K_M	16.559 GB	26.896B	Not recorded	—	73.041 GB	Get weights
Qwen3.5-27BQ4_K_M · Pareto choice	16.741 GB	27.781B	86.1	5.143	72.859 GB	Get weights Score source
Qwen3.8-27B Uncensored (JonathanColetti)Q4_K_M	16.811 GB	27.321B	Not recorded	—	72.789 GB	Get weights
InnoSpark3.1-27B-1008Q4_K_M	16.811 GB	27.781B	Not recorded	—	72.789 GB	Get weights
Qwen3.6-27BQ4_K_M · Pareto choice	16.817 GB	27.781B	86.2	5.126	72.783 GB	Get weights Score source
Holo4-27BQ4_K_M	16.875 GB	27.357B	Not recorded	—	72.725 GB	Get weights
gemma-4-26B-A4B-itQ4_K_M	16.948 GB	25.806B	82.6	4.874	72.652 GB	Get weights Score source
Swift 1.5 Qwen3.8-27BQ4_K_M	17.442 GB	27.321B	Not recorded	—	72.158 GB	Get weights
ThinkingCap Qwen3.8-27BQ4_K_M	17.442 GB	27.321B	Not recorded	—	72.158 GB	Get weights
Swift-Qwen3.8-27bQ4_K_M	17.442 GB	27.781B	84.95	4.87	72.158 GB	Get weights Score source
Cloudflare Clef 27BQ4_K_MDecision model. llama.cpp uses /v1/systemone; text generation is unavailable. Image input requires current model support plus a compatible multimodal projector. See the Clef comparison for sources and checks.	17.652 GB	26.896B	Not recorded	—	71.948 GB	Get weights
GLM-4.7-FlashQ4_K_M	18.312 GB	31.221B	Not recorded	—	71.288 GB	Get weights
gemma-4-31B-itQ4_K_M	18.324 GB	31.273B	85.2	4.65	71.276 GB	Get weights Score source
Qwen3-30B-A3B-Instruct-2507Q4_K_M	18.557 GB	30.532B	78.4	4.225	71.043 GB	Get weights Score source
Qwen3-30B-A3B-Thinking-2507Q4_K_M	18.557 GB	30.532B	80.9	4.36	71.043 GB	Get weights Score source
Qwen3.8-27B EfficientThink SFT-SimPOQ4-LynnStyle	19.352 GB	26.896B	Not recorded	—	70.248 GB	Get weights
d3 27BQ4_K_M	19.404 GB	26.896B	Not recorded	—	70.196 GB	Get weights
DeepSeek-R1-Distill-Qwen-32BQ4_K_M	19.851 GB	32.764B	Not recorded	—	69.749 GB	Get weights
Xing4.0-29B-A4B-GGUFIQ4_NLPublished tests use very long contexts; this weight budget does not reproduce those settings.	20.104 GB	31.215B	Not recorded	—	69.496 GB	Get weights
Xing4.0 29B-A4BIQ4_NLPublished tests use very long contexts; this weight budget does not reproduce those settings.	20.104 GB	31.215B	Not recorded	—	69.496 GB	Get weights
MiniCPM-V 4.7 35B-A3BQ4_K_M	21.166 GB	34.66B	Not recorded	—	68.434 GB	Get weights
Occamy-1.0 35B-A3BQ4_K_M	21.167 GB	34.661B	Not recorded	—	68.433 GB	Get weights
APUS Reflex 35B-A3BQ4_K_M	21.167 GB	34.661B	Not recorded	—	68.433 GB	Get weights
Holo4-35B-A3BQ4_K_M	21.298 GB	35.107B	Not recorded	—	68.302 GB	Get weights
Qwen3.8-35B-A3B-DistillQ4_K_M	21.713 GB	35.505B	Not recorded	—	67.887 GB	Get weights
Seed-OSS-36B-InstructQ4_K_M	21.762 GB	36.151B	82.7	3.8	67.838 GB	Get weights Score source
Ornith-1.5-35B-A3BQ4_K_M	21.864 GB	35.952B	Not recorded	—	67.736 GB	Get weights
Qwen3.6-35B-A3BQ4_K_M	22.135 GB	35.952B	85.2	3.849	67.465 GB	Get weights Score source
Qwen3.5-35B-A3BQ4_K_M	22.285 GB	35.952B	85.3	3.828	67.315 GB	Get weights Score source
Tiel-Coder 35B-A3BUD-Q4_K_XL	22.36 GB	34.661B	73.7	3.296	67.24 GB	Get weights Score source
K2 Horizon MoVA 36B-A4BQ4_K_M	22.368 GB	37.445B	Not recorded	—	67.232 GB	Get weights
NVIDIA-Nemotron-3-Nano-30B-A3BQ4_K_M	24.574 GB	31.578B	78.3	3.186	65.026 GB	Get weights Score source
NVIDIA-Nemotron-3.5-Lightning-30B-A3BQ4_K_M	25.477 GB	32.913B	81.94	3.216	64.123 GB	Get weights Score source
Midnight-Miqu-70B-v1.5Q4_K_M	41.423 GB	68.977B	31.39	0.758	48.177 GB	Get weights Score source
Aleph Alpha Kolibri-1Q4_K_M	47.454 GB	78.103B	80	1.686	42.146 GB	Get weights Score source
Xor 1.1BF16	70.214 GB	35.107B	Not recorded	—	19.386 GB	Get weights

Additional candidates with 64 GB dedicated VRAM

The 51.2 GB VRAM weight budget admits these additional files. They do not pass this page’s more cautious system-memory budget.

MMLU-Pro (where recorded) · GB = 1 billion bytes. Published scores may use different settings or precision.
Model / quantization	Weights	Parameters	Score / 100	Score / GB	Budget left	Sources
Aleph Alpha Kolibri-1Q4_K_M · Pareto choice	47.454 GB	78.103B	80	1.686	3.746 GB	Get weights Score source

Computers near this memory tier

These are memory candidates, not tested model/runtime matches. Dedicated VRAM is used for GPU-fit planning; system RAM is not treated as VRAM.

NVIDIA Jetson AGX Orin 64GB Developer Kit · 64 GB unified memory
GMKtec EVO-X3 · 128 GB unified memory
GMKtec EVO-X2 · 64 GB unified memory
ASUS Ascent GX10 · 128 GB unified memory
BOSGAME M5 · 128 GB unified memory
Beelink GTR9 Pro 395 · 128 GB unified memory
GMKtec EVO-T1 · 64 GB system RAM
ACEMAGIC M1A PRO+ · 128 GB unified memory
NVIDIA DGX Spark · 128 GB unified memory
AMD Ryzen AI Halo · 128 GB unified memory
msecore Ryzen AI 9 HX 370 Workstation · 64 GB system RAM
MINISFORUM MS-S1 Max · 128 GB unified memory
MSI EdgeXpert 11SUS · 128 GB unified memory
Acer Veriton GN100-UD11 · 128 GB unified memory
GMKtec EVO-X2 · 128GB · 128 GB unified memory
MINISFORUM AI X1 Pro · 96GB · 96 GB system RAM
GMKtec EVO-T2S · 64GB · 64 GB system RAM
MINISFORUM N5 MAX · 64GB · 64 GB unified memory
NIMO AI Mini PC · 128GB · 128 GB unified memory
Moginsok MG01 · 128GB · 128 GB unified memory

Compare all hardware

How recommendations are calculated

We compare one benchmark at a time. A Pareto choice has no other scored model with both an equal or smaller weight file and an equal or higher score, with at least one strict improvement. The first choice is the frontier point nearest the chart’s top-left corner, using the same rounded memory scale and 0–100 score scale as the main chart. Other frontier choices follow, then other scored models by score.

Score per GB is shown for context; it does not choose the champion. Changing the budget or dataset can change the result. Scores are published point estimates, often for the source checkpoint rather than the selected quantization. They do not establish statistically significant differences or equal evaluation conditions.

RAM, VRAM and unified memory

Use one memory pool. System RAM is used for CPU inference; a dedicated GPU has separate VRAM. Do not add RAM and VRAM as if they were one GPU pool. Partial CPU offload needs a supported backend and can change speed. Unified memory is shared by CPU, GPU and the operating system; software can limit how much the GPU can allocate.

Our initial weight budget reserves 30% of system RAM or unified memory, with a minimum of 2 GB. For VRAM it reserves 20%, with a minimum of 1 GB. These are adjustable planning assumptions, not measured requirements. Context cache, runtime buffers, other programs, multimodal files and concurrent users can need more. A file that passes the budget check is a candidate, not a confirmed hardware-compatible configuration.

Check the model’s runtime and quantization support before downloading. See the Transformers cache documentation and llama.cpp backend documentation. GB uses decimal bytes; operating systems can report GiB instead.

Related memory tiers and tasks

All model records · Memory planning guide