How much VRAM does Inkling need?

IQ1_M is exactly 226,224,289,216 bytes (210.69 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Inkling's KV cache?

8.25 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

Is Inkling a mixture-of-experts model?

Yes — 256 experts with 6 routed per token. Every expert must be resident in memory, but only the routed ones are read per token, so its memory requirement and its speed behave very differently.

thinkingmachines · vision language · mixture of experts

Inkling

thinkingmachines/Inkling

Inkling at IQ1_M is exactly 226,224,289,216 bytes (210.69 GiB / 226.22 GB) — an effective 1.900 bits per weight, not the nominal 1. Its KV cache at 32K is 8.25 GiB.

From the file· summed from 6 file(s)From the file· KV per layer

Parameters

952B

total, not active

Architecture

inkling

66 layers

Context

—

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
IQ1_M6 shards	210.69 GiB	226,224,289,216	1.900	—	AtomicChat
UD-IQ1_S7 shards	251.61 GiB	270,163,818,071	2.269	—	unsloth
UD-IQ1_M7 shards	265.46 GiB	285,036,820,087	2.394	—	unsloth
MXFP413 shards	478.42 GiB	513,695,069,824	4.315	—	AtomicChat
Q8_019 shards	797.92 GiB	856,755,583,751	7.197	—	unsloth
Q8_025 shards	937.42 GiB	1,006,542,568,544	8.455	—	AtomicChat
BF1641 shards	1764.18 GiB	1,894,278,563,239	15.912	—	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	1.03 GiB	1.03 GiB	—	66 / 0 / 0
8,192	2.06 GiB	2.06 GiB	—	66 / 0 / 0
16,384	4.13 GiB	4.13 GiB	—	66 / 0 / 0
32,768	8.25 GiB	8.25 GiB	—	66 / 0 / 0
65,536	16.50 GiB	16.50 GiB	—	66 / 0 / 0
131,072	33.00 GiB	33.00 GiB	—	66 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP vs Qwen3.5-122B-A10B vs Kimi-K2.7-Code vs Step-3.7-Flash vs Qwen3.5-397B-A17B vs Ornith-1.0-397B All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts IQ1_M at roughly 498.92 GiB. The real file is 210.69 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

6144

Vocab

201,024

Sliding window

none

SWA period

—

MLA

Experts

256

Experts per token

use_sliding_window

—

Questions people ask

How much VRAM does Inkling need?: IQ1_M is exactly 226,224,289,216 bytes (210.69 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Inkling's KV cache?: 8.25 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Is Inkling a mixture-of-experts model?: Yes — 256 experts, 6 routed per token. Every expert must be resident, but only the routed ones are read per token, which is why its memory requirement and its speed behave very differently.
Which quantization of Inkling should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.