TechCompare LogoTechCompare

How much VRAM does Inkling 975B (MoE) need at Q4_K_M? Thinking Machines 1M hybrid

Inkling 975B at Q4_K_M is a 653 GB resident deployment for 1M context, or about 77 GB with active-expert offload.

Inkling 975B at Q4_K_M with native 1M context needs about 653 GB of VRAM with all experts resident. Thinking Machines' flagship MoE activates 41B parameters per token from a 975B total pool. The hybrid global and sliding-window attention keeps the key-value cache under 48 GB at the full 1M context. Active-expert offload drops the resident footprint to roughly 77 GB.

By TechCompare · Updated

Total VRAM required
653 GB
Inkling 975B (MoE) at Q4_K_M
Weights
546 GB
975B params
KV cache
47.4 GB
1024K tokens, FP16 KV

Calculator

Estimated VRAM required

653 GB

975B params at Q4_K_M, 1,048,576 token context, batch 1, inference.

Weights
546 GB
KV cache
47.4 GB
Overhead
59.3 GB

Estimate accuracy: Weights within ~2%. KV cache within ~5% for standard GQA models, ~10% for MLA or hybrid state estimates. Real VRAM may vary with framework (vLLM vs llama.cpp vs Transformers), Flash Attention, and driver overhead.

Sliding-window attention applied: This model caps 5 of every 6 layers at a 512-token window. KV cache estimate is 83% smaller than naive full-attention math at this context length.

Hardware that fits

Apple entries show installed unified memory separately from a model budget based on an 8 GB reserve. That budget is a planning heuristic, not a fixed hardware limit.

No single GPU in our catalog has enough memory. Multi-GPU or CPU offload required.

How this is calculated

The 975B parameter pool requires 546 GB of weights at Q4_K_M. The 66-layer architecture uses group-query attention with variable KV heads: 8 in the 11 global layers, 16 in the 55 sliding-window layers. The sliding window caps 55 of 66 layers at a 512-token window with only 11 layers doing full global attention. The result is a 47 GB key-value cache at the full 1M context. Activation and software overhead adds about 59 GB. In active-only mode the resident weights shrink to 23 GB, which drops total usage to 77 GB.

Verdict

The 653 GB resident total breaks into 546 GB of Q4 weights, 47 GB of FP16 KV cache at the full 1M context (variable KV heads: 8 in 11 global layers, 16 in 55 sliding-window layers capped at a 512-token window), and 59 GB of overhead. Hardware-wise resident mode needs nine H100 80GB cards or five H200 141GB cards with NVLink. Active-only loading shrinks the weights to about 23 GB by routing cold experts through system RAM, while the 47 GB cache stays, dropping the total to roughly 77 GB. That fits a single 80 GB pro card with headroom for batch, or a 96 GB unified-memory workstation. The 512-token sliding window on 55 of 66 layers is what keeps the cache tractable at full context, which is why active-only is viable here at all. At 256K the active-only total drops to about 54 GB with even more batch headroom.

More Inkling scenarios

DeepSeek V4 Pro 1.6T (MoE) at Q4_K_M
DeepSeek V4 Pro 1.6T at Q4_K_M with the full 1M-token context needs about 1012 GB of VRAM with every expert resident - this is the real shape of the model and the number to plan a deployment against.
View details ➜
Llama 4 Scout (17B/109B) at Q4_K_M
At Q4_K_M and 10,485,760 tokens, this model estimates 2335 GB with all 109B parameters resident.
View details ➜
gpt-oss 20B (MoE) at Q4_K_M
gpt-oss 20B at Q4_K_M with native 128K context needs about 19.4 GB of VRAM with all experts resident, dropping to roughly 9.3 GB with active-only weight loading.
View details ➜

Frequently asked questions

Why does Inkling use variable KV heads across layers?
The 11 global-attention layers use 8 KV heads while the 55 sliding-window layers use 16 KV heads. The sliding-window layers are capped at a 512-token window, so their extra KV heads cost little at long context. The global layers dominate the cache at full context, so the effective count used in sizing is 8.
Can I run Inkling on a single GPU?
In active-only mode, yes. The 77 GB active-only total fits on an 80GB H100 or A100. Resident mode needs 653 GB across nine 80GB cards. Capping context to 256K drops active-only to about 54 GB, which leaves comfortable headroom on a single 80GB card.
Can I run Inkling 975B on a single GPU?
No. The 975B pool at Q4 is roughly 546 GB of weights, exceeding every consumer GPU and most datacenter GPUs. Active-only loading could fit the resident expert subset on a single 80 GB card, but the cold-expert bandwidth penalty makes per-token throughput low. For most users, the hosted API from Thinking Machines Lab or OpenRouter partners is the practical option.
When does the hosted API beat self-hosting for a 975B-class MoE?
Almost always, until your monthly volume reaches the multi-hundred-million-token range with strict data residency requirements. The self-hosting bill at this scale is a small cluster, ops staffing, and utilization risk, while the API is a line-item you can dial to zero. The self-host argument is strongest for burst-proof capacity planning and air-gapped data, not for everyday cost.