Inkling Small GPU requirements: VRAM and cheapest GPU
Inkling Small has 266B parameters, published in BF16. It is a mixture-of-experts model: 6 of its 256 experts work on each token, about 1.7B parameters, so it generates at roughly the speed of a model that size. Every expert still has to sit in GPU memory, which is why the memory figures below are for the full model. It is released under apache-2.0, which allows commercial use.
- Parameters
- 266B
- Smallest footprint (INT4)
- about 156 GB
- Cheapest GPU per hour (FP16)
- $2.39
- Type
- Vision
thinkingmachines/Inkling-Small
- Parameters 266B
- Active per token 1.7B
- Type Vision
- Task image-text-to-text
- Licence apache-2.0
- Downloads 916.1K
- Likes 460
Calculating for Inkling Small
More settings
Inkling Small needs about 588 GB of GPU memory
Inference (vLLM), FP16 / BF16, 8K tokens context, one request at a time.
Where the memory goes
- Weights 532 GB
- KV cache 1.4 GB
- Runtime overhead 54 GB
- Total 588 GB
GPU counts keep 10% of each card free for spikes. Plan extra if you expect long prompts or many users at once.
| Rank | Configuration | Per GPU / hr | GPUs needed | Total / hr | Total / month (24/7) | Provider |
|---|---|---|---|---|---|---|
| 1 | 4 × MI300X 192GBCheapestCDNA 3 · HBM3 | $2.39 | 4 | $9.56 | $6,979 | RunPod (opens in a new tab)4x MI300X |
| 2 | 7 × RTX PRO 6000 96GBBlackwell · GDDR7 | $1.80 | 7sold as 8 GPUs; you need 7 | $14.40 | $10,512 | Nebius (opens in a new tab)8x RTX PRO 6000 96GB |
| 3 | 7 × H100 NVL 94GBHopper · HBM3 | $3.19 | 7sold as 8 GPUs; you need 7 | $25.52 | $18,630 | RunPod (opens in a new tab)8x H100 NVL |
| 4 | 5 × H200 141GBHopper · HBM3e | $3.79 | 5sold as 8 GPUs; you need 5 | $30.32 | $22,134 | RunPod (opens in a new tab)8x H200 NVL |
| 5 | 4 × B200 180GBBlackwell · HBM3e | $7.79 | 4 | $31.16 | $22,747 | Lambda (opens in a new tab)4x B200 SXM6 |
| 6 | 3 × B300 288GBBlackwell Ultra · HBM3e | $8.99 | 3sold as 4 GPUs; you need 3 | $35.96 | $26,251 | RunPod (opens in a new tab)4x B300 |
| 7 | 4 × GB200 186GBBlackwell · HBM3e | $10.50 | 4 | $42.00 | $30,660 | CoreWeave (opens in a new tab)GB200 NVL72 |
| More than 8: needs several nodes: H100 80GB, A100 80GB, L40S 48GB, RTX 6000 Ada 48GB, RTX A6000 48GB, A40 48GB, A100 40GB, RTX 5090 32GB, RTX 4090 24GB, RTX 3090 24GB, L4 24GB, A10 24GB, T4 16GB, V100 16GB | ||||||
| Provider | Instance | Per GPU / hr | GPUs needed | Total / hr | Total / month (24/7) |
|---|---|---|---|---|---|
| RunPod (opens in a new tab)Cheapest | 4x MI300Xus | $2.39 | 4 × MI300X 192GB | about $9.56 | about $6,979 |
| Nebius (opens in a new tab) | 8x RTX PRO 6000 96GBeu/us | $1.80 | 7 × RTX PRO 6000 96GB sold as 8 GPUs; you need 7 | about $14.40 | about $10,512 |
| Vast.ai (opens in a new tab) | 8x RTX PRO 6000 S (median of 4 offers)worldwide marketplace | $2.23 | 7 × RTX PRO 6000 96GB sold as 8 GPUs; you need 7 | about $17.87 | about $13,045 |
| CoreWeave (opens in a new tab) | RTX PRO 6000 Blackwell Server Edition (High Memory)us | $2.50 | 7 × RTX PRO 6000 96GB sold as 8 GPUs; you need 7 | about $20.00 | about $14,600 |
| Lambda (opens in a new tab) | 4x B200 SXM6us | $7.79 | 4 × B200 180GB | about $31.16 | about $22,747 |
| Vultr (opens in a new tab) | vbm-256c-2048gb-8-mi300x-gpu (bare metal)us | $3.99 | 4 × MI300X 192GB sold as 8 GPUs; you need 4 | about $31.92 | about $23,302 |
| AWS (opens in a new tab) | g7e.48xlargeus-east-1 | $4.14 | 7 × RTX PRO 6000 96GB sold as 8 GPUs; you need 7 | about $33.14 | about $24,195 |
| Google Cloud (opens in a new tab) | g4-standard-384us-central1 | $4.50 | 7 × RTX PRO 6000 96GB sold as 8 GPUs; you need 7 | about $36.00 | about $26,280 |
| Azure (opens in a new tab) | ND128isr_NDR_GB200_v6eastus | $27.04 | 4 × GB200 186GB | about $108.16 | about $78,957 |
Prices collected Nov 16, 2026. Every cost is approximate: on-demand, US dollars, mostly US regions (marketplaces such as Vast.ai and Lium are worldwide).
Memory by precision and workload
| Precision | Inference | LoRA fine-tune | QLoRA fine-tune | Full fine-tune |
|---|---|---|---|---|
| FP16 / BF16 | 588 GB | 684 GB | – | 4,686 GB |
| INT8 | 310 GB | – | – | – |
| INT4 | 156 GB | – | 253 GB | – |
Cheapest GPU by precision
Inference with an 8K context and one request at a time.
FP16 / BF16
about 588 GB · 8K tokens
The weights as published. Nothing lost, but the largest footprint.
4 × MI300X 192GB
RunPod (opens in a new tab) · 4x MI300X
- Per GPU / hr
- $2.39
- GPUs needed
- 4
- Total / hr
- about $9.56
- per month, 24/7, approx.
- $6,979
INT8
about 310 GB · 8K tokens
Weights at half the size. Most chat and RAG work shows no visible difference.
8 × A40 48GB
RunPod (opens in a new tab) · 8x A40
- Per GPU / hr
- $0.59
- GPUs needed
- 8
- Total / hr
- about $4.72
- per month, 24/7, approx.
- $3,446
INT4
about 156 GB · 8K tokens
Weights at a quarter of the size. Small drops show up first in maths and code.
4 × A40 48GB
RunPod (opens in a new tab) · 4x A40
- Per GPU / hr
- $0.59
- GPUs needed
- 4
- Total / hr
- about $2.36
- per month, 24/7, approx.
- $1,723
Inference vs fine-tuning
Inference (FP16)
1×
of the memory inference needs
588 GB
4 × MI300X 192GB · about $9.56
LoRA fine-tune
1.2×
of the memory inference needs
684 GB
4 × MI300X 192GB · about $9.56
QLoRA fine-tune
0.4×
of the memory inference needs
253 GB
6 × A40 48GB · about $4.72
Full fine-tune
8×
of the memory inference needs
4,686 GB
more than one 8-GPU node
Serving it at FP16 fits on 4 × MI300X 192GB (RunPod). LoRA needs 4 × MI300X 192GB (RunPod), QLoRA 6 × A40 48GB (RunPod) and a full fine-tune more than one 8-GPU node. Training keeps activations and optimiser state in memory as well as the weights, which is why it so often needs more GPUs than serving the same model.
Will it run on a Mac or a single RTX 4090?
RTX 4090 (24 GB)
Not on its own. Even at INT4 it needs about 156 GB, more than the card’s 24 GB. Two or more cards, or a bigger data-centre GPU, will do it.
Mac with Apple silicon
Yes, with enough memory. At INT4 it needs about 156 GB, so a Mac with 256 GB or more of unified memory (macOS lets the GPU use about three quarters of it). Expect slower generation than on a data-centre GPU.
Similar models
Open models of a similar size.
- MiniMax-M3.1-FlashVision220.5B · about 487 GB FP16 · 150K ↓GPU needs
- GLM-5.3-FlashVision321.3B · about 708 GB FP16 · 6.4M ↓GPU needs
- Command A+Vision218.8B · about 483 GB FP16 · 8.9K ↓GPU needs
- DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder (GGUF)Vision196.6B · about 437 GB FP16 · 525.1K ↓GPU needs
- Qwen3.8-Flash-NextVision180B · about 397 GB FP16 · 2.7M ↓GPU needs
- Llama 4 MaverickVision401.6B · about 886 GB FP16 · 14.4K ↓GPU needs
Inkling Small: common questions
How much VRAM does Inkling Small need?
About 588 GB at FP16, 310 GB at INT8 and 156 GB at INT4, with an 8K context and one request at a time. Longer contexts and more parallel requests add KV cache on top.
What is the cheapest GPU to run Inkling Small?
4 × MI300X 192GB (about $2.39 per GPU / hour) on RunPod, at about $9.56 an hour for FP16 / BF16 inference with an 8K context. Prices collected Nov 16, 2026; they change often.
Can I fine-tune Inkling Small on a single GPU?
Not on one GPU. Even QLoRA needs about 253 GB, more than an 80 GB GPU holds. LoRA needs about 684 GB and a full fine-tune about 4,686 GB.
How are these numbers calculated?
Weights are parameters times bytes per parameter (2 at FP16, 1 at INT8, 0.5 at INT4). The KV cache comes from the model’s layers, KV heads, head size and context length. We add about 10% plus 1 GB for the runtime and fill each GPU to 90%. Prices are on-demand list prices in US regions, collected from each provider’s public pricing.
Can Inkling Small run on a consumer GPU like the RTX 4090?
Not on its own. Even at INT4 it needs about 156 GB, more than the card’s 24 GB. Two or more cards, or a bigger data-centre GPU, will do it.
Does Inkling Small fit on a single H100 80GB?
No. Even at INT4 it needs about 156 GB, so you need several GPUs.
How many GPUs do I need for Inkling Small?
At FP16 with an 8K context: 9 × 80 GB, 14 × 48 GB or 28 × 24 GB GPUs. At INT4 you need about a quarter of the memory for the weights.
How much VRAM does Inkling Small need at INT8?
About 310 GB: 279 GB for the weights, 1.4 GB for an 8K KV cache and 29 GB of runtime overhead.
How much does it cost to run Inkling Small 24/7 for a month?
About $6,979 on the cheapest setup we found (4 × MI300X 192GB (about $2.39 per GPU / hour), $9.56 an hour for 730 hours). The cheapest on AWS, Google Cloud or Azure is about $24,195 a month.
FP16 vs INT8 vs INT4: which should I use?
FP16 keeps full quality and needs the most memory (about 588 GB here). INT8 halves the weights with little loss for most chat work. INT4 (AWQ, GPTQ or GGUF Q4) cuts them to a quarter, about 156 GB in total, with a small drop in quality that shows most in maths and code.
Can Inkling Small run on a Mac?
Yes, with enough memory. At INT4 it needs about 156 GB, so a Mac with 256 GB or more of unified memory (macOS lets the GPU use about three quarters of it). Expect slower generation than on a data-centre GPU.
vLLM or llama.cpp?
vLLM is built for serving many requests at once on NVIDIA and AMD GPUs and reserves most of the GPU memory for KV cache. llama.cpp runs GGUF files on almost anything, including Macs and CPUs, and suits one user or a small team. Memory needs are similar for the same precision.
Prices collected Nov 16, 2026. Every cost is approximate: on-demand, US dollars, mostly US regions (marketplaces such as Vast.ai and Lium are worldwide).
How we calculate
- Weights. Parameters times bytes per parameter: 2 at FP16, 1 at INT8, 0.5 at INT4, plus about 5% for the quantisation scales at INT8 and INT4.
- KV cache. 2 × layers × KV heads × head size × context × requests × 2 bytes, from the model’s config.json. Where a setting is missing we estimate it and say so.
- Overhead. About 10% plus 1 GB for the runtime, in line with what vLLM reserves.
- Fine-tuning. LoRA keeps 16-bit weights and adds a small adapter, its optimiser and activations. QLoRA does the same over 4-bit weights. A full fine-tune needs about 16 bytes per parameter with mixed-precision Adam (weights, gradients and optimiser states), plus activations.
- GPUs. We fill each GPU to 90% of its memory and stop at 8 per node. Costs are on-demand list prices in US regions; marketplace clouds show the median offer.