Thinking Machines Lab, United States · Released July 15, 2026
Inkling has 952.4B parameters (41B active per token). In BF16 its weights alone take 1764 GiB; quantized to 4-bit, about 533 GiB. Each 32K-token request adds 1.59 GiB of KV cache. Cheapest way to run it today: 8 × NVIDIA A100 80GB at $5.04 per hour. It reads text, images and audio, calls tools and lets you set its reasoning effort.
Read every day from Hugging Face. Split files are added together; vision projectors are left out.
How many GPUs to run Inkling
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
GPU
NVFP4 experts + BF16 (native)
BF16 / FP16
8-bit
4-bit
GeForce RTX 4090 24 GB
Not on this GPU
96 (several machines)
48 (several machines)
32 (several machines)
GeForce RTX 5090 32 GB
Not on this GPU
64 (several machines)
32 (several machines)
24 (several machines)
RTX PRO 6000 Blackwell 96 GB
Not on this GPU
24 (several machines)
16 (several machines)
6
Mac M5 Max 128 GB 96 GB usable
Not on this GPU
24 (several machines)
16 (several machines)
6
Ryzen AI Max+ 395 128 GB 96 GB usable
Not on this GPU
24 (several machines)
16 (several machines)
6
DGX Spark (GB10) 128 GB
Not on this GPU
16 (several machines)
8
5
NVIDIA H100 80 GB
16 (several servers)
32 (several servers)
16 (several servers)
8
NVIDIA H200 141 GB
5
16 (several servers)
8
5
NVIDIA B200 (HGX) 180 GB
4
16 (several servers)
6
4
8-bit: FP8 with vLLM (INT8 on GPUs without FP8), Q8_0 with llama.cpp. 4-bit: AWQ / GPTQ with vLLM, Q4_K_M with llama.cpp, both counted at the Q4_K_M size. Not on this GPU: the official NVFP4 experts + BF16 weights do not run there (llama.cpp needs a GGUF version, and vLLM needs a GPU with FP8 support).
Cheapest way to run Inkling today
For each GPU: the fewest cards that fit Inkling with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-10-30).
On-demand prices from AWS, Google Cloud, Azure, RunPod, Vast.ai, Oracle Cloud, OVHcloud, Scaleway, Verda, Akash and UpCloud, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
Reasoning. Adjustable reasoning effort. Reasoning tokens take context and KV cache like any other token: leave room for them.
KV cache: how memory grows with context
Architecture: MoE 256 experts + 2 shared · 11 global layers (8 KV × 128) + 55 sliding-window layers (512 tokens, 16 KV × 128) · short convolutions · image and audio input. Its KV cache is 10.4× smaller than if every layer used full attention, because only 55 of its 66 layers keep a KV cache; the others hold a small fixed-size state. Only 41B parameters are active per token, but all 952B must sit in memory: VRAM depends on the total, speed on the active part.
Context per request
BF16 cache
FP8 cache
8K tokens
0.56 GiB
0.28 GiB
32K tokens
1.59 GiB
0.79 GiB
128K tokens
5.71 GiB
2.86 GiB
1M tokens
44.2 GiB
22.1 GiB
Plus 8.25 MiB of recurrent state per request, whatever the context.
Inkling VRAM FAQ
How much VRAM does Inkling need?
In BF16 the weights alone take 1764 GiB (1894 GB). In FP8 that is 884 GiB, and about 533 GiB with 4-bit quantization (Q4_K_M); the official NVFP4 experts + BF16 checkpoint is 542 GiB. Each request then adds KV cache: 0.56 GiB at 8K tokens and 1.59 GiB at 32K (BF16 cache).
Can Inkling run on a single RTX 4090 (24 GB)?
No. Even in 4-bit it needs 32 RTX 4090s, spread over several machines. In 4-bit it needs 8 H100 GPUs.
Can Inkling run on an RTX 5090 (32 GB)?
No. Even in 4-bit it needs 24 RTX 5090s, spread over several machines.
Can Inkling run on a Mac?
No single Mac holds it, even in 4-bit: it needs 533 GiB with an 8K-token context, and the largest Mac (512 GB) gives its GPU about 384 GB.
Can Inkling run on a DGX Spark or a Ryzen AI Max+ 395 (128 GB)?
No. Even in 4-bit it needs 5 DGX Spark units or 6 Ryzen AI Max+ 395 machines (128 GB), linked over the network.
What is the cheapest way to run Inkling?
On 2026-10-30, the cheapest on-demand setup is 8 × NVIDIA A100 80GB with 4-bit weights on Vast.ai, at $5.04 per hour (about $3,679.20 per month). Next: 6 × RTX PRO 6000 Blackwell (96 GB) with Q4_K_M weights on Vast.ai, at $8.00 per hour. Sized for one 8K-token request; prices are checked every hour.
How many H100 GPUs does Inkling need?
With one request and an 8K-token context: 32 in BF16, 16 in FP8 and 8 in 4-bit (vLLM, 90% of memory usable). Above 8 GPUs, that means several servers. Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
Can Inkling read images?
Yes. Inkling reads text, images and audio. Images add tokens to the context, so they also add KV cache.
Does Inkling support tool calling?
Yes, Inkling supports tool (function) calling: it can ask your application to run a function and use the result.
Does Inkling have a thinking mode?
Adjustable reasoning effort. Reasoning tokens count toward the context and the KV cache.
Inkling VRAM badge
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/inkling.json.
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.
Thinking Machines Lab, United States · Released July 15, 2026