What can you run with 24GB of RAM?
With 24GB RAM you can run 55 of the 99 open-source models we track — up to about 30B parameters at Q4 quantization. The most popular that fits is gemma-4-26B-A4B-it, needing 20 GB and generating around ~31 tok/s on typical hardware of this size.
Typical machines: MacBook Air M4 (24GB) · MacBook Pro M4 Pro (base) · Mac mini M4 Pro. After the operating system takes its share, about 21 GB is free for a model.
Having the memory is not the same as having the speed
24GB RAM decides whether a model loads. How fast it answers is decided by memory bandwidth — how quickly your machine can read the model, which it must do once for every word it writes. Machines with 24GB RAM span 68–307 GB/s, so the fastest is about 4.5× quicker than the slowest with identical capacity. The table quotes a M4; pick yours below to see the difference. Why bandwidth and not the processor →
Models that run on 24GB RAM in 2026
| Model↕ | Size↕ | Memory↑ | Est. speed↕ | Downloads↕ | Feels like |
|---|---|---|---|---|---|
| granite-embedding-97m-multilingual-r2Ibm | 97M | 655 MBRuns well | ~740 tok/s | 294,744 | Fast |
| granite-embedding-311m-multilingual-r2Ibm | 312M | 748 MBRuns well | ~415 tok/s | 47,167 | Fast |
| LFM2.5-Encoder-230MLiquid Ai | 230M | 793 MBRuns well | ~562 tok/s | 15,930 | Fast |
| LFM2.5-230MLiquid Ai | 230M | 808 MBRuns well | ~508 tok/s | 59,937 | Fast |
| ERNIE-4.5-0.3B-PTBaidu AI | 361M | 853 MBRuns well | ~324 tok/s | 21,323 | Fast |
| nemotron-3.5-asr-streaming-0.6bNVIDIA | 638M | 1.1 GBRuns well | ~157 tok/s | 1,037,937 | Fast |
| Qwen3.5-0.8BAlibaba / Qwen | 873M | 1.3 GBRuns well | ~146 tok/s | 2,990,867 | Fast |
| LFM2.5-1.2B-InstructLiquid Ai | 1.2B | 1.4 GBRuns well | ~107 tok/s | 580,233 | Fast |
| LFM2.5-VL-1.6BLiquid Ai | 1.6B | 1.4 GBRuns well | ~107 tok/s | 66,900 | Fast |
| Nemotron-3-Embed-1B-BF16NVIDIA | 1.1B | 1.5 GBRuns well | ~113 tok/s | 446,251 | Fast |
| cohere-transcribe-arabic-07-2026Cohere | 2.1B | 1.9 GBRuns well | ~63 tok/s | 47,612 | Fast |
| Hy-MT2-1.8BTencent | 2B | 1.9 GBRuns well | ~69 tok/s | 150,417 | Fast |
| Qwen3.5-2BAlibaba / Qwen | 2.3B | 2.0 GBRuns well | ~61 tok/s | 2,468,876 | Fast |
| granite-speech-4.1-2bIbm | 2.3B | 2.4 GBRuns well | ~50 tok/s | 445,508 | Fast |
| LFM2.5-Audio-1.5BLiquid Ai | 1.5B | 2.6 GBRuns well | ~43 tok/s | 1,266 | Fast |
| Unlimited-OCRBaidu AI | 3.3B | 2.7 GBRuns well | ~40 tok/s | 2,536,284 | Fast |
| LocateAnything-3BNVIDIA | 3.8B | 2.8 GBRuns well | ~37 tok/s | 671,404 | Fast |
| granite-4.1-3bIbm | 3.4B | 3.0 GBRuns well | ~37 tok/s | 350,238 | Fast |
| Ministral-3-3B-Instruct-2512Mistral AI | 3.8B | 3.1 GBRuns well | ~36 tok/s | 544,419 | Fast |
| Nemotron-Labs-Diffusion-3B-BaseNVIDIA | 3.8B | 3.3 GBRuns well | ~34 tok/s | 79,309 | Fast |
| Cosmos3-EdgeNVIDIA | 3.9B | 3.3 GBRuns well | ~33 tok/s | 46,719 | Fast |
| Nemotron-3.5-Content-SafetyNVIDIA | 4.3B | 3.6 GBRuns well | ~31 tok/s | 112,409 | Fast |
| gemma-4-E2B-itGoogle DeepMind | 5.1B | 3.8 GBRuns well | ~25 tok/s | 4,116,770 | Fast |
| Qwen3.5-4BAlibaba / Qwen | 4.7B | 3.8 GBRuns well | ~28 tok/s | 6,199,203 | Fast |
| Fara1.5-4BMicrosoft | 4.5B | 4.0 GBRuns well | ~27 tok/s | 4,587 | Fast |
| NVIDIA-Nemotron-3-Nano-4B-BF16NVIDIA | 4B | 4.1 GBRuns well | ~27 tok/s | 712,701 | Fast |
| Fara-7BMicrosoft | 8.3B | 5.5 GBRuns well | ~17 tok/s | 2,695 | Comfortable |
| Hy-MT2-7BTencent | 8B | 5.7 GBRuns well | ~17 tok/s | 35,238 | Comfortable |
| LFM2.5-8B-A1BLiquid Ai · MoE | 8.5B | 5.9 GBRuns well | ~128 tok/s1B active | 170,261 | Fast |
| Nemotron-3-Embed-8B-BF16NVIDIA | 8B | 5.9 GBRuns well | ~16 tok/s | 122,309 | Comfortable |
| gemma-4-E4B-itGoogle DeepMind | 8B | 6.2 GBRuns well | ~15 tok/s | 5,863,029 | Comfortable |
| Nemotron-Labs-Diffusion-8B-BaseNVIDIA | 8.5B | 6.2 GBRuns well | ~15 tok/s | 440,687 | Comfortable |
| Ministral-3-8B-Instruct-2512Mistral AI | 8.9B | 6.3 GBRuns well | ~15 tok/s | 119,379 | Comfortable |
| granite-4.1-8bIbm | 8.8B | 6.3 GBRuns well | ~15 tok/s | 3,991,168 | Comfortable |
| Ornith-1.0-9BDeepreinforce | 9B | 6.7 GBRuns well | ~14 tok/s | 2,250,422 | Comfortable |
| Qwen3.5-9BAlibaba / Qwen | 9.7B | 6.8 GBRuns well | ~14 tok/s | 11,767,971 | Comfortable |
| GLM-4.6V-FlashZhipu AI | 10B | 7.0 GBRuns well | ~13 tok/s | 282,562 | Comfortable |
| Fara1.5-9BMicrosoft | 9.4B | 7.0 GBRuns well | ~13 tok/s | 13,751 | Comfortable |
| Olmo-3-7B-InstructAi2 | 7.3B | 7.2 GBRuns well | ~17 tok/s | 406,776 | Comfortable |
| gemma-4-12B-itGoogle DeepMind | 12B | 9.4 GBRuns well | ~11 tok/s | 2,987,745 | Comfortable |
| Ministral-3-14B-Instruct-2512Mistral AI | 14B | 9.7 GBRuns well | ~9.5 tok/s | 198,169 | Comfortable |
| Nemotron-Labs-Diffusion-14BNVIDIA | 14B | 9.7 GBRuns well | ~9.5 tok/s | 7,571 | Comfortable |
| Cosmos3-NanoNVIDIA | 16B | 11 GBRuns well | ~8.2 tok/s | 342,949 | Comfortable |
| gpt-oss-20bOpenAI · MoE | 22B | 13 GBRuns well | ~34 tok/s4.3B active | 8,250,813 | Fast |
| ERNIE-4.5-21B-A3B-PTBaidu AI · MoE | 22B | 15 GBRuns well | ~42 tok/s3B active | 44,971 | Fast |
| Voxtral-Small-24B-2507Mistral AI | 24B | 16 GBRuns well | ~5.5 tok/s | 279,195 | Slow |
| Magistral-Small-2506Mistral AI | 24B | 16 GBRuns well | ~5.4 tok/s | 39,823 | Slow |
| Qwen3.5-27BAlibaba / Qwen | 28B | 19 GBTight | ~4.7 tok/s | 2,637,888 | Slow |
| diffusiongemma-26B-A4B-itGoogle DeepMind · MoE | 26B | 19 GBTight | ~30 tok/s4B active | 2,008,287 | Fast |
| gemma-4-26B-A4B-itGoogle DeepMind · MoE | 27B | 20 GBTight | ~31 tok/s4B active | 12,150,722 | Fast |
| ERNIE-4.5-VL-28B-A3B-PTBaidu AI · MoE | 29B | 20 GBTight | ~43 tok/s3B active | 218,879 | Fast |
| Qwen3.6-27BAlibaba / Qwen | 28B | 20 GBTight | ~4.6 tok/s | 6,572,759 | Slow |
| granite-4.1-30bIbm | 29B | 20 GBTight | ~4.5 tok/s | 20,171 | Slow |
| Fara1.5-27BMicrosoft | 27B | 20 GBTight | ~4.4 tok/s | 2,938 | Slow |
| Hy-MT2-30B-A3BTencent · MoE | 30B | 20 GBTight | ~43 tok/s3B active | 164,966 | Fast |
Sorted by memory, smallest first — click any column to change it.
Step up to 32GB RAM and 8 more models come within reach
Including gemma-4-31B-it (33B), Qwen3.6-35B-A3B (36B) and Ornith-1.0-35B (35B) — around ~3.2 tok/s on a typical machine that size.
See everything 32GB RAM runs →All figures at Q4_K_M quantization, 4K context. “Memory” includes the model, its context cache and runtime overhead. Speed is calculated, not benchmarked: 120 GB/s × 0.65 ÷ bytes read per token, assuming a M4. Scale it by your own machine's bandwidth from the list above. Mixture-of-experts models read only their active parameters per token, which is why some large ones outrun smaller dense models.
Image generators that also fit
These make pictures rather than text. Their memory use scales with the resolution you generate at, not conversation length, so treat these figures as a floor rather than a ceiling.
What you'd need more memory for
The next models up, and what they ask for.