d1-3B requirements — can you run it?
~2.7B parameters · 4 quantizations measured · updated 2026-10-07 · source: LiquidAI/d1-3B-GGUF
# llama.cpp (auto-downloads the GGUF) llama-cli -hf LiquidAI/d1-3B-GGUF:Q8_0 # Ollama (pulls straight from Hugging Face) ollama run hf.co/LiquidAI/d1-3B-GGUF:Q8_0 # LM Studio: search "LiquidAI/d1-3B-GGUF" in the model browser
Commands load the exact quant measured on this page.
Every quantization: real file size vs what you actually need
| Quant | File (measured) | VRAM/RAM needed (8K ctx) | Runs on (first fit) |
|---|---|---|---|
| Q4_K_M | 1.7 GB | 3.5 GB | RTX 3060 12GB |
| Q8_0 | 2.9 GB | 4.9 GB | RTX 3060 12GB |
| BF16 | 5.4 GB | 8.0 GB | RTX 3060 12GB |
| F16 | 5.4 GB | 8.0 GB | RTX 3060 12GB |
KV cache uses a flat +20% context/overhead rule because this repo does not publish its config; exact numbers may differ for long contexts. MoE models keep all experts in memory — total size counts, not just active params. Longer context grows the KV cache — use the selector above.
Can I run d1-3B on my GPU? (Q8_0 recommended quant)
| Hardware | Memory | Verdict | Est. speed |
|---|---|---|---|
| RTX 3060 12GB | 12 GB | RUNS | ~94 tok/s |
| RTX 4060 Ti 16GB | 16 GB | RUNS | ~75 tok/s |
| RTX 3090 24GB | 24 GB | RUNS | ~244 tok/s |
| RTX 4090 24GB | 24 GB | RUNS | ~263 tok/s |
| RTX 5090 32GB | 32 GB | RUNS | ~468 tok/s |
| RTX PRO 6000 96GB | 96 GB | RUNS | ~468 tok/s |
| Mac 16GB unified | 16 GB | RUNS | ~26 tok/s |
| Mac 32GB unified | 32 GB | RUNS | ~39 tok/s |
| Mac 64GB unified | 64 GB | RUNS | ~71 tok/s |
| Mac 128GB unified | 128 GB | RUNS | ~104 tok/s |
| Mac 256GB unified | 256 GB | RUNS | ~142 tok/s |
| Mac 512GB unified | 512 GB | RUNS | ~214 tok/s |
| 32GB system RAM (CPU) | 32 GB | RUNS | ~13 tok/s |
| 64GB system RAM (CPU) | 64 GB | RUNS | ~13 tok/s |
| 128GB system RAM (CPU) | 128 GB | RUNS | ~21 tok/s |
| 256GB system RAM (CPU) | 256 GB | RUNS | ~21 tok/s |
| 1TB server RAM (CPU) | 1024 GB | RUNS | ~52 tok/s |
| 2TB server RAM (CPU) | 2048 GB | RUNS | ~52 tok/s |
Speed = decode tok/s estimated from memory bandwidth ÷ measured file size (× MoE active share). Real numbers vary ±30% by runtime and settings.
How much VRAM does d1-3B need?
The honest answer: 4.9 GB at the Q8_0 quant with 8K context — that is the measured
file size (2.9 GB) plus KV cache and runtime overhead. The smallest published build needs
3.5 GB; the lossless one needs 8.0 GB.
Cheapest hardware that runs it: Used RTX 3070 8GB (~$220, used market) — see which GPU to buy.
Can I run it with CPU offload?
Yes, if combined RAM+VRAM ≥ file size — but generation speed is limited by memory bandwidth. Expect roughly: DDR5 dual channel ~10-30 tok/s for small MoE actives, single digits for big ones, 0.1-1 tok/s when experts page from disk. CPU-only is fine for batch jobs, painful for chat.
Why do other sites show different numbers?
Most pages compute weights from a formula (params × bits/8). We use the actual file sizes published in GGUF repos, which include the embedding table, unquantized tensors, and container overhead — that is why our numbers can differ from a naive calculation.
FAQ
How much VRAM does d1-3B need?
The measured answer: 4.9 GB at the Q8_0 quant with 8K context. The smallest published build needs 3.5 GB; the lossless (F16/BF16) build needs 8.0 GB. These are real GGUF file sizes from Hugging Face, not formula estimates.
Can I run d1-3B on a 24GB GPU (RTX 3090/4090)?
Yes — at Q8_0 (4.9 GB). Quants that fit 24GB: Q4_K_M, Q8_0, BF16, F16.
How fast will d1-3B run?
Decode speed is memory-bandwidth-bound. At Q8_0: roughly 263 tok/s on an RTX 4090, 244 tok/s on an RTX 3090, 104 tok/s on an M-series Mac with 128GB, and 13.0 tok/s on CPU with dual-channel DDR4 — estimates from measured file size, MoE active-parameter share and memory bandwidth.
Can I run d1-3B on CPU with system RAM?
Yes, if your usable RAM is at least the file size. At Q8_0 you need ~4.9 GB of RAM. CPU generation is memory-bandwidth-bound: expect single-digit tok/s for large MoE models, 10-30 tok/s for small active params, and below 1 tok/s when experts page from disk.
Why do VRAM numbers for d1-3B differ between sites?
Most sites compute weights from a formula (params x bits/8). ModelFit uses the actual GGUF file sizes published on Hugging Face, which include the embedding table, unquantized tensors, and container overhead — so our numbers reflect what you really download and load.