What can you run with 8GB of RAM?
With 8GB RAM you can run 51 of the 129 open-source models we track — up to about 16B parameters at Q4 quantization. The most popular that fits is Qwen3.5-4B, needing 3.8 GB and generating around ~12 tok/s on typical hardware of this size.
Typical machines: Base MacBook Air (M1/M2) · Budget Windows laptops · Entry-level desktops. After the operating system takes its share, about 6.4 GB is free for a model.
Having the memory is not the same as having the speed
8GB 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 8GB RAM span 51–102 GB/s, so the fastest is about 2.0× quicker than the slowest with identical capacity. The table quotes a DDR4-3200; pick yours below to see the difference. Why bandwidth and not the processor →
Models that run on 8GB RAM in 2026
| Model↕ | Size↕ | Memory↑ | Est. speed↕ | Downloads↕ | Feels like |
|---|---|---|---|---|---|
| granite-embedding-97m-multilingual-r2Ibm | 97M | 603 MBRuns well | ~563 tok/s | 97,079 | Fast |
| LFM2.5-Encoder-230MLiquid Ai | 230M | 793 MBRuns well | ~239 tok/s | 13,432 | Fast |
| LFM2.5-230MLiquid Ai | 230M | 808 MBRuns well | ~216 tok/s | 55,999 | Fast |
| granite-embedding-311m-multilingual-r2Ibm | 312M | 821 MBRuns well | ~131 tok/s | 64,530 | Fast |
| granite-speech-5.0-470m-turboctcIbm | 473M | 857 MBRuns well | ~116 tok/s | 11,294 | Fast |
| harrier-oss-v1-270mMicrosoft | 268M | 865 MBRuns well | ~131 tok/s | 727,222 | Fast |
| LFM2.5-Encoder-350M-Policy-LinterLiquid Ai | 355M | 885 MBRuns well | ~155 tok/s | 9,957 | Fast |
| LFM2.5-VL-450MLiquid Ai | 449M | 900 MBRuns well | ~145 tok/s | 67,396 | Fast |
| nemotron-3.5-asr-streaming-0.6bNVIDIA | 638M | 1.1 GBRuns well | ~67 tok/s | 917,259 | Fast |
| Qwen3.5-0.8BAlibaba / Qwen | 873M | 1.3 GBRuns well | ~62 tok/s | 2,881,964 | Fast |
| privacy-filterOpenAI · MoE | 1.4B | 1.4 GBRuns well | ~196 tok/s280M active | 456,933 | Fast |
| LFM2.5-1.2B-InstructLiquid Ai | 1.2B | 1.4 GBRuns well | ~45 tok/s | 366,847 | Fast |
| LFM2.5-VL-1.6BLiquid Ai | 1.6B | 1.4 GBRuns well | ~45 tok/s | 248,110 | Fast |
| Nemotron-3-Embed-1B-BF16NVIDIA | 1.1B | 1.5 GBRuns well | ~48 tok/s | 386,914 | Fast |
| Hy-MT2-1.8BTencent | 2B | 1.9 GBRuns well | ~29 tok/s | 24,475 | Fast |
| MiniMax-Music3MiniMax | 2.4B | 2.1 GBRuns well | ~24 tok/s | 23,217 | Comfortable |
| Qwen3.5-2BAlibaba / Qwen | 2.3B | 2.1 GBRuns well | ~24 tok/s | 3,097,484 | Comfortable |
| cohere-transcribe-arabic-07-2026Cohere | 2.1B | 2.3 GBRuns well | ~21 tok/s | 51,893 | Comfortable |
| cohere-transcribe-03-2026Cohere | 2.1B | 2.3 GBRuns well | ~21 tok/s | 607,964 | Comfortable |
| granite-speech-4.1-2bIbm | 2.3B | 2.4 GBRuns well | ~21 tok/s | 237,181 | Comfortable |
| LFM2.5-2.6BLiquid Ai | 2.7B | 2.5 GBRuns well | ~20 tok/s | 133,677 | Comfortable |
| LFM2.5-VL-3BLiquid Ai | 3.1B | 2.5 GBRuns well | ~20 tok/s | 24,282 | Comfortable |
| granite-4.0-1b-speechIbm | 2.3B | 2.5 GBRuns well | ~21 tok/s | 74,644 | Comfortable |
| LFM2.5-Audio-1.5BLiquid Ai | 1.5B | 2.6 GBRuns well | ~18 tok/s | 1,702 | Comfortable |
| Unlimited-OCRBaidu AI | 3.3B | 2.7 GBRuns well | ~17 tok/s | 3,008,635 | Comfortable |
| Nemotron-Labs-Diffusion-3BNVIDIA | 3B | 2.8 GBRuns well | ~18 tok/s | 35,414 | Comfortable |
| LocateAnything-3BNVIDIA | 3.8B | 2.8 GBRuns well | ~16 tok/s | 95,054 | Comfortable |
| granite-4.1-3bIbm | 3.4B | 2.9 GBRuns well | ~16 tok/s | 122,076 | Comfortable |
| granite-4.2-3bIbm | 3.7B | 3.1 GBRuns well | ~15 tok/s | 14,073 | Comfortable |
| granite-4.0-3b-visionIbm | 4B | 3.1 GBRuns well | ~15 tok/s | 1,069 | Comfortable |
| Shieldstral-1.0-3BMistral AI | 3.8B | 3.1 GBRuns well | ~15 tok/s | 20,895 | Comfortable |
| Ministral-3-3B-Instruct-2512Mistral AI | 3.8B | 3.1 GBRuns well | ~15 tok/s | 544,419 | Comfortable |
| Nemotron-Labs-Diffusion-3B-BaseNVIDIA | 3.8B | 3.3 GBRuns well | ~14 tok/s | 19,213 | Comfortable |
| Cosmos3-EdgeNVIDIA | 3.9B | 3.3 GBRuns well | ~14 tok/s | 1,325,763 | Comfortable |
| Nemotron-3.5-Content-SafetyNVIDIA | 4.3B | 3.6 GBRuns well | ~13 tok/s | 134,547 | Comfortable |
| gemma-4-E2B-itGoogle DeepMind | 5.1B | 3.8 GBRuns well | ~11 tok/s | 3,235,641 | Comfortable |
| Qwen3.5-4BAlibaba / Qwen | 4.7B | 3.8 GBRuns well | ~12 tok/s | 7,681,585 | Comfortable |
| Fara1.5-4BMicrosoft | 4.5B | 4.0 GBRuns well | ~11 tok/s | 2,786 | Comfortable |
| MolmoWeb-4BAi2 | 4.9B | 4.1 GBRuns well | ~11 tok/s | 2,152 | Comfortable |
| NVIDIA-Nemotron-3-Nano-4B-BF16NVIDIA | 4B | 4.1 GBRuns well | ~12 tok/s | 712,701 | Comfortable |
| Fara-7BMicrosoft | 8.3B | 5.5 GBRuns well | ~7.1 tok/s | 1,593 | Slow |
| Hy-MT2-7BTencent | 8B | 5.7 GBTight | ~7.2 tok/s | 12,004 | Slow |
| LFM2.5-8B-A1BLiquid Ai · MoE | 8.5B | 5.9 GBTight | ~54 tok/s1B active | 96,675 | Fast |
| Nemotron-3-Embed-8B-BF16NVIDIA | 8B | 5.9 GBTight | ~6.9 tok/s | 81,804 | Slow |
| Nemotron-Labs-Diffusion-8BNVIDIA | 8B | 5.9 GBTight | ~6.9 tok/s | 163,233 | Slow |
| Cosmos3-NanoNVIDIA | 16B | 6.2 GBTight | ~6.6 tok/s | 244,209 | Slow |
| gemma-4-E4B-itGoogle DeepMind | 8B | 6.2 GBTight | ~6.2 tok/s | 4,740,694 | Slow |
| Nemotron-Labs-Diffusion-8B-BaseNVIDIA | 8.5B | 6.2 GBTight | ~6.5 tok/s | 280,417 | Slow |
| Ministral-3-8B-Instruct-2512Mistral AI | 8.9B | 6.3 GBTight | ~6.4 tok/s | 119,379 | Slow |
| granite-4.1-8bIbm | 8.8B | 6.3 GBTight | ~6.5 tok/s | 1,441,995 | Slow |
| MolmoWeb-8BAi2 | 8.7B | 6.4 GBTight | ~6.3 tok/s | 1,679 | Slow |
Sorted by memory, smallest first — click any column to change it.
Step up to 16GB RAM and 9 more models come within reach
Including Qwen3.5-9B (9.7B), gemma-4-12B-it (12B) and Ornith-1.0-9B (9B) — around ~10 tok/s on a typical machine that size.
See everything 16GB 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: 51 GB/s × 0.65 ÷ bytes read per token, assuming a DDR4-3200. 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.