MODEL DIRECTORY
Local LLM models: sizes, benchmarks and sources
Browse 129 models with up to 80B total parameters or a selected weight set up to 40 GB. Each page includes the selected weight file, published scores and test settings.
Find models for your computer · Compare by memory and task
Scores from different benchmarks are not combined. Use the interactive ranking to compare models on one test.
| Model | Weight size | Total parameters | Quantization | Published tests |
|---|---|---|---|---|
| Afrique Qwen v2 | 2.709 GB | 4.206B | Q4_K_M | AfriMMLU |
| Agents A1 4B | 2.709 GB | 4.539B | Q4_K_M | LiveCodeBench v6, IFEval · reported, SciCode |
| Ai2 AstaBrief 8B | 5.028 GB | 8.191B | Q4_K_M | ScholarQA-CS2 · test average |
| Aleph Alpha Kolibri-1 | 47.454 GB | 78.103B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026, SWE-bench Verified, Terminal-Bench 2.1 |
| APUS Reflex 35B-A3B | 21.167 GB | 34.661B | Q4_K_M | APUS Reflex · Frozen80 Q4 |
| APUS Reflex 4B | 2.709 GB | 4.206B | Q4_K_M | APUS Reflex 4B · Frozen80 Q4 |
| APUS Reflex 9B | 5.629 GB | 9.41B | Q4_K_M | APUS Reflex 9B · Frozen80 Q4 |
| Arm-optimized Qwen3.5-2B | 1.274 GB | 1.882B | Q4_K_M | MMLU-Redux |
| Bonsai 27B (1-bit) | 3.803 GB | 26.896B | Q1_0 | HumanEval+, IFEval · reported, GSM8K, AIME 2025, AIME 2026, MBPP+, MATH-500, MMLU-Redux, LiveCodeBench · version unspecified, BFCL v3, IFBench · prompt loose |
| Clary 2 Cyber Heavy | 2.497 GB | 4.022B | Q4_K_M | SecEval |
| Cloudflare Clef 27B | 17.652 GB | 26.896B | Q4_K_M | Decision Index 0.2.1 · BFCL |
| Cloudflare Clef-flash 9B | 6.041 GB | 9.076B | Q4_K_M | Decision Index 0.2.1 · BFCL |
| ColGrep-Agent-2B | 1.561 GB | 2.517B | Q4_K_M | SWE-bench Lite localization · File F1, SWE-bench Lite localization · Line F1, Multi-SWE-bench Flash localization · File F1, Multi-SWE-bench Flash localization · Line F1 |
| Curious 2B | 1.215 GB | 1.882B | Q4_0 | CuriousLM · phone turns passed |
| d3 27B | 19.404 GB | 26.896B | Q4_K_M | Decision Index 0.3.1 · Composite |
| d3-edge 0.6B | 0.806 GB | 0.748B | Q8_0 | Decision Index 0.3.1 · Composite |
| DeepSeek-R1-Distill-Qwen-32B | 19.851 GB | 32.764B | Q4_K_M | GPQA Diamond, MATH-500, LiveCodeBench · version unspecified, AIME 2024 · pass@1 |
| DFM Mimir | 1.913 GB | 1.787B | Q8_0 | HumanEval, HellaSwag · reported, MMLU, GSM8K |
| DuoNeural Hermes v2 | 5.156 GB | 8.468B | Q4_K_M | HumanEval, HumanEval+, MBPP+ |
| DuoNeural Hermes v3 | 5.156 GB | 8.468B | Q4_K_M | HumanEval, HumanEval+, MBPP+ |
| Empero Qwen3.8 2B | 1.312 GB | 2.274B | Q4_K_M | GSM8K, MMLU CoT · ~1,700 questions |
| Empero Qwen3.8 4B | 2.783 GB | 4.66B | Q4_K_M | GSM8K, MMLU CoT · ~1,700 questions |
| Empero Qwen3.8 9B | 5.78 GB | 9.653B | Q4_K_M | GSM8K, MMLU CoT · ~1,700 questions |
| FrogNano-4B-2609 | 2.798 GB | 4.326B | Q4_K_M | SWE-bench Verified · avg@3 |
| Gemma 4 E2B | 3.107 GB | 5.123B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026 |
| Gemma 4 E4B | 4.977 GB | 7.996B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026 |
| gemma-4-12B-it | 7.122 GB | 11.96B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026 |
| gemma-4-26B-A4B-it | 16.948 GB | 25.806B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026 |
| gemma-4-31B-it | 18.324 GB | 31.273B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026 |
| GLM-4.7-Flash | 18.312 GB | 31.221B | Q4_K_M | LiveCodeBench v6, AIME 2025, SWE-bench Verified, GPQA · variant unspecified |
| GLM-4.7-Flash-REAP-23B-A3B | 14.114 GB | 22.996B | Q4_K_M | HumanEval, HumanEval+ |
| Gmcoder 9B | 9.786 GB | 8.954B | Q8_0 | HumanEval, HumanEval+ |
| Granite 4.2 3B | 2.244 GB | 3.66B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2025, GPQA · variant unspecified |
| Granite 4.2 8B | 5.348 GB | 8.792B | Q4_K_M | MMLU-Pro, LiveCodeBench v6, AIME 2025, GPQA · variant unspecified |
| Granite concision edit | 2.673 GB | 1.335B | F16 | GSM8K · OVR 50 |
| grug 3B | 2.575 GB | 4.17B | Q4_K_M | GSM8K · grug 80 |
| H2O-Lightning 4B Native | 4.482 GB | 4.206B | Q8_0 | JEVBench · public 231 |
| Holo4-27B | 16.875 GB | 27.357B | Q4_K_M | OSWorld · Holo harness, OSWorld 2.0 · partial score, AutomationBench · Holo held-out 120 |
| Holo4-35B-A3B | 21.298 GB | 35.107B | Q4_K_M | OSWorld · Holo harness, OSWorld 2.0 · partial score, AutomationBench · Holo held-out 120 |
| Humanizer 12B RLRt2 | 12.67 GB | 11.907B | Q8_0 | Humanizer fact fidelity · EN 420 |
| Index Translate 2B | 1.312 GB | 2.274B | Q4_K_M | FLORES · COMET-22 (126k / 420) |
| Index-Translate-9B | 5.78 GB | 9.653B | Q4_K_M | FLORES COMET-22 · Index-Translate, WMT26 Judge · Index-Translate, instTrans IFscore · Index-Translate |
| InnoSpark3.1-27B-1008 | 16.811 GB | 27.781B | Q4_K_M | LongTutor · XES3G5M Gold memory |
| Josev-2B | 2.045 GB | 1.912B | Q8_0 | Josev-2B Q8_0 · decisions-de validation |
| K2 Horizon 0.9B | 0.666 GB | 1.078B | Q4_K_M | GPQA Diamond, LiveCodeBench v6, HumanEval+, AIME 2025, AIME 2026, HMMT February 2026, BFCL v4, MBPP+ |
| K2 Horizon 3.7B | 3.157 GB | 5.058B | Q4_K_M | GPQA Diamond, HMMT February 2026, BFCL v4, SciCode, SWE-bench Verified |
| K2 Horizon 7B | 5.592 GB | 8.999B | Q4_K_M | HMMT February 2026, SciCode, SWE-bench Verified |
| K2 Horizon MoVA 36B-A4B | 22.368 GB | 37.445B | Q4_K_M | GPQA Diamond, SciCode, Terminal-Bench 2.1 |
| Kernel Logic 1.5 | 4.683 GB | 7.616B | Q4_K_M | MMLU-Pro, HumanEval, GSM8K |
| LFM2.5 2.6B | 1.674 GB | 2.697B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2025, BFCL v4 |
| LFM2.5 230M | 0.153 GB | 0.23B | Q4_K_M | MMLU-Pro, GPQA Diamond, IFEval · reported, BFCL v4 |
| LFM2.5-24B-A2B | 14.526 GB | 23.972B | Q4_K_M | BFCL v4 |
| LightOnOCR-3 0.8B | 0.529 GB | 0.853B | Q4_K_M | olmOCR-Bench · LightOnOCR-3 |
| Limite 1B - Violetto | 2.071 GB | 1B | BF16 | AIME 2026, BeyondAIME |
| Ling 3.0 tiny | 4.824 GB | 7.893B | Q4_K_M | GPQA Diamond, HMMT February 2026, BFCL v4, SciCode |
| Liquid AI d1-3B | 2.875 GB | 3.123B | Q8_0 | Decision Index 0.2.1 · Composite |
| Liquid AI d1-omni-600M | 0.407 GB | 0.587B | Q8_0 | Decision Index 0.2.1 · Composite |
| Lumma-Fev-27B | 16.547 GB | 26.085B | Q4_K_M | Lumma-Fev-27B · held-out decisions |
| Mellum2.1 12B A2.5B Thinking | 8.071 GB | 12.15B | Q4_K_M | LiveCodeBench v6, BFCL v4, SWE-bench Verified, Terminal-Bench 2.1 |
| Midnight-Miqu-70B-v1.5 | 41.423 GB | 68.977B | Q4_K_M | MMLU-Pro, GPQA · variant unspecified |
| MiMo-V2.6-Distill-Qwen-9B | 5.841 GB | 8.954B | Q4_K_M | SWE-bench Verified · avg@3 |
| MiniCPM-V 4.7 1B | 0.672 GB | 1.006B | Q4_K_M | MMBench · reported, OCRBench · reported, ScreenSpot · reported |
| MiniCPM-V 4.7 35B-A3B | 21.166 GB | 34.66B | Q4_K_M | MMBench · reported, OCRBench · reported, ScreenSpot · reported |
| MiniCPM5 2B | 1.561 GB | 2.517B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, IFEval · reported, AIME 2025, AIME 2026 |
| MiniCPM5 abliterated v2 | 1.561 GB | 2.517B | Q4_K_M | MMLU-Pro · 350 × 2 seeds, MATH-500 |
| Mistral-Nemo-Instruct-2407 | 7.477 GB | 12.248B | Q4_K_M | HellaSwag · reported, MMLU |
| Mitsuba-ComfyUI-27B | 5.998 GB | 27.321B | PTQ1_0 | Mitsuba 10-axis suite · v1.18 |
| Nanbeige 4.2 3B | 2.684 GB | 4.17B | Q4_K_M | GPQA Diamond, LiveCodeBench v6, HMMT February 2026, SciCode, SWE-bench Verified |
| NeoHorse 1 4B | 2.709 GB | 4.206B | Q4_K_M | LiveCodeBench v6, HumanEval, IFEval · reported, BFCL v4 |
| Noema 1.5 2B | 1.274 GB | 1.882B | Q4_K_M | GPQA Diamond, HumanEval+, IFEval · prompt strict, MMLU-Pro · 11,332 holdout |
| Noema 2B | 2.012 GB | 1.882B | Q8_0 | MMLU-Pro, HumanEval, HumanEval+, MBPP+ |
| NVIDIA-Nemotron-3-Nano-30B-A3B | 24.574 GB | 31.578B | Q4_K_M | MMLU-Pro, AIME 2025, BFCL v4, GPQA · variant unspecified |
| NVIDIA-Nemotron-3.5-Lightning-30B-A3B | 25.477 GB | 32.913B | Q4_K_M | MMLU-Pro, GPQA Diamond, SciCode, SWE-bench Verified |
| Occamy-1.0 35B-A3B | 21.167 GB | 34.661B | Q4_K_M | IFEval · reported, BFCL v4 |
| OneJev 1.5 0.8B | 0.672 GB | 1.007B | Q4_K_M | Decision Index 0.3 · Public |
| OpenDecider-small-td | 2.497 GB | 4.022B | Q4_K_M | typed-decisions · OpenDecider 2,000 |
| OpenJev 27B | 16.547 GB | 26.896B | Q4_K_M | OpenJev GGUF validation · 1,789 |
| OpenMobile-2 27B | 16.547 GB | 26.896B | Q4_K_M | AndroidWorld Pass@1 · OpenMobile-2, MobileWorld GUI · OpenMobile-2, MobileGym SR · OpenMobile-2, MobileGym++ Hybrid · OpenMobile-2, SPA-Bench L3 · OpenMobile-2 |
| OpenMobile-2 9B | 5.629 GB | 8.954B | Q4_K_M | AndroidWorld Pass@1 · OpenMobile-2, MobileWorld GUI · OpenMobile-2, MobileGym SR · OpenMobile-2, MobileGym++ Hybrid · OpenMobile-2, SPA-Bench L3 · OpenMobile-2 |
| Ornith-1.5-35B-A3B | 21.864 GB | 35.952B | Q4_K_M | GPQA Diamond, SWE-bench Verified |
| Ornith-1.5-9B | 5.78 GB | 9.197B | Q4_K_M | GPQA Diamond, SWE-bench Verified |
| PicoLM 80M | 0.18 GB | 0.09B | F16 | HellaSwag · reported |
| PicoLM v2 81M | 0.193 GB | 0.096B | F16 | HellaSwag · reported |
| PicoLM v2.1 81M | 0.193 GB | 0.096B | F16 | HellaSwag · reported |
| Quill 1 Preview | 0.398 GB | 0.494B | Q4_K_M | HumanEval · Quill 50 |
| Quinto 0.1 4B | 4.482 GB | 4.206B | Q8_0 | IFEval · prompt strict |
| Qwen3-30B-A3B-Instruct-2507 | 18.557 GB | 30.532B | Q4_K_M | MMLU-Pro, LiveCodeBench v6, IFEval · reported, AIME 2025, GPQA · variant unspecified, MMLU-Redux, BFCL v3 |
| Qwen3-30B-A3B-Thinking-2507 | 18.557 GB | 30.532B | Q4_K_M | MMLU-Pro, LiveCodeBench v6, IFEval · reported, AIME 2025, GPQA · variant unspecified, MMLU-Redux, BFCL v3 |
| Qwen3.5-27B | 16.741 GB | 27.781B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, IFEval · reported, BFCL v4, SWE-bench Verified, MMLU-Redux |
| Qwen3.5-35B-A3B | 22.285 GB | 35.952B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, IFEval · reported, BFCL v4, SWE-bench Verified, MMLU-Redux |
| Qwen3.5-4B | 2.741 GB | 4.206B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, IFEval · reported, BFCL v4, MMLU-Redux |
| Qwen3.5-9B | 5.681 GB | 9.653B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, IFEval · reported, BFCL v4, MMLU-Redux |
| Qwen3.6-27B | 16.817 GB | 27.781B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026, HMMT February 2026, SWE-bench Verified, MMLU-Redux |
| Qwen3.6-35B-A3B | 22.135 GB | 35.952B | Q4_K_M | MMLU-Pro, LiveCodeBench v6, AIME 2026, HMMT February 2026, SWE-bench Verified, GPQA · variant unspecified, MMLU-Redux |
| Qwen3.8 27B Humanlike Chat 2.0 | 16.559 GB | 26.896B | Q4_K_M | IFEval · prompt strict |
| Qwen3.8-27B | 16.464 GB | 27.781B | Q4_K_M | GPQA Diamond, LiveCodeBench v6 |
| Qwen3.8-27B EfficientThink SFT-SimPO | 19.352 GB | 26.896B | Q4-LynnStyle | GPQA Diamond, MMLU · nerkyor 500-question suite, LiveCodeBench · nerkyor 100-task suite |
| Qwen3.8-27B Uncensored (JonathanColetti) | 16.811 GB | 27.321B | Q4_K_M | MMLU · 0-shot reported, HellaSwag · 0-shot reported |
| Qwen3.8-27B-GSQ-RCO-GGUF | 10.094 GB | 26.896B | IQ3_XXS | GPQA Diamond, LiveCodeBench v6, AIME 2025 |
| Qwen3.8-35B-A3B-Distill | 21.713 GB | 35.505B | Q4_K_M | MMLU |
| Qwythos-9B Claude Mythos 5 1M | 5.629 GB | 8.954B | Q4_K_M | GSM8K · Qwythos 100, GPQA Diamond · Qwythos 100 |
| RapidChat 9B | 5.629 GB | 9.41B | Q4_K_M | τ²-bench airline test · pass^1 |
| RefinedToolCall v5 | 2.538 GB | 3.086B | Q6_K | AIME 2024 · pass@8 |
| Satella-RAG 0.8B | 0.593 GB | 0.773B | Q5_K_M | Satella RAG · overall reference F1 |
| Seed-OSS-36B-Instruct | 21.762 GB | 36.151B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, IFEval · reported, MMLU, AIME 2025, SWE-bench Verified |
| Spark X2.5 1.7B | 1.107 GB | 1.708B | Q4_K_M | IFEval · reported, AIME 2026, HMMT February 2026, BFCL v4, SciCode, SWE-bench Verified, GPQA · variant unspecified |
| Spark X2.5 4B | 2.6 GB | 4.112B | Q4_K_M | IFEval · reported, AIME 2026, HMMT February 2026, BFCL v4, SciCode, SWE-bench Verified, GPQA · variant unspecified |
| StartLux-Decision-v1-2B | 1.274 GB | 1.882B | Q4_K_M | JEVBench · public 231 |
| Swift 1.5 Qwen3.8-27B | 17.442 GB | 27.321B | Q4_K_M | GPQA Diamond, LiveCodeBench v6, AIME 2026 |
| Swift Bonsai 2 27B (PQ2_0) | 7.206 GB | 26.896B | PQ2_0 | IFBench · prompt loose |
| Swift-Qwen3.8-27b | 17.442 GB | 27.781B | Q4_K_M | MMLU-Pro, GPQA Diamond, LiveCodeBench v6, AIME 2026 |
| Switchboard Router 0.5B v0.1 | 0.994 GB | 0.494B | F16 | Switchboard v0.1 · Tier accuracy |
| Switchboard Router 4B | 4.28 GB | 4.022B | Q8_0 | Switchboard v0.1 · Tier accuracy |
| Ternary Bonsai 2 27B | 5.947 GB | 26.896B | PTQ1_0 | HumanEval+, IFEval · reported, GSM8K, AIME 2025, AIME 2026, MBPP+, MATH-500, MMLU-Redux, LiveCodeBench · version unspecified, BFCL v3, IFBench · prompt loose |
| Ternary Bonsai 2 27B Abliterated (BoldingBuilds) | 5.947 GB | 26.896B | PTQ1_0 | HumanEval, MMLU · 0-shot reported |
| Ternary Bonsai 27B | 7.165 GB | 26.896B | PQ2_0 | HumanEval+, IFEval · reported, GSM8K, AIME 2025, AIME 2026, MBPP+, MATH-500, MMLU-Redux, LiveCodeBench · version unspecified, BFCL v3, IFBench · prompt loose |
| ThakiCloud RAG-Gate-27B | 16.547 GB | 27.321B | Q4_K_M | RAG-Gate · blind action accuracy |
| ThinkingCap Qwen3.8-27B | 17.442 GB | 27.321B | Q4_K_M | GPQA Diamond, MMLU-Pro · ThinkingCap 1,500, IFBench · ThinkingCap 300 × 2, AA-LCR · ThinkingCap 100, RealWorldQA · ThinkingCap 765 × 2 |
| Tiel-Coder 35B-A3B | 22.36 GB | 34.661B | UD-Q4_K_XL | MMLU-Pro, SWE-bench-Live · Tiel 25, Claw-Eval · Tiel 114 |
| TwIL LM3 | 1.915 GB | 3.075B | Q4_K_M | GSM8K · TwIL 300 |
| Udgam 384M Chat | 0.238 GB | 0.384B | Q4_K_M | BFCL v3 simple · 150 |
| UltraCoder L3 | 5.156 GB | 8.468B | Q4_K_M | HumanEval, HumanEval+, MBPP+ |
| Vela Lumen v1.1 | 0.143 GB | 0.048B | F16 | HellaSwag · Vela 5,000 |
| Wald-Q4B v2.1 | 2.709 GB | 4.66B | Q4_K_M | JEVBench · public 231 |
| webAI TwIL-LM3-Pro | 2.244 GB | 3.66B | Q4_K_M | MATH-500 |
| Winnow 12B | 12.67 GB | 11.96B | Q8_0 | JevBench public subset · Winnow 231, Kev v9 clean · Winnow 1,046 |
| Xing4.0 29B-A4B | 20.104 GB | 31.215B | IQ4_NL | AIME 2026, SWE-bench Verified, Terminal-Bench 2.1 |
| Xing4.0-29B-A4B-GGUF | 20.104 GB | 31.215B | IQ4_NL | AIME 2026, SWE-bench Verified |
| Xor 1.1 | 70.214 GB | 35.107B | BF16 | JEVBench · public 231 |
File size is not the total RAM requirement. Context cache and runtime memory need extra space. See the local LLM hardware comparison for memory types and system limits.