Evidence: benchmark results vendor-reported
Inkling 2 Small — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
The second-generation Inkling: a 594B mixture-of-experts with 17B active per token (8 of 384 experts), 48 layers of which 40 use a 1,024-token sliding window and 8 are global, image input and a 1M-token context. It was trained natively in NVFP4, and the 393.5 GB NVFP4 checkpoint is the one Thinking Machines recommends; an MXFP8 build (622.8 GB) is about 1-2 points better on evals. Apache-2.0; official vLLM and SGLang recipes.
Inkling 2 Small needs about 394 GB of VRAM at NVFP4 + MXFP8 (mixed) — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Includes the KV cache at 8K context, so it reads higher than the headline figure.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| NVFP4 + MXFP8 (mixed) 5.30 bpw | 394.8 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Inkling 2 Small VRAM calculator →
Get Inkling 2 Small running
Inkling 2 Small is not practical on a single consumer GPU: it needs about 394.3 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Inkling 2 Small
Ollama lists Inkling 2 Small only as the cloud tag inkling-2-small:cloud, which runs on Ollama's servers, not your machine. Locally, Thinking Machines documents vLLM and SGLang on a multi-GPU server.
Specifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 594 Billion (17B active)
- Context window
- 1,048,576
- Architecture
- Mixture-of-Experts (384 experts, 8 active), 40 sliding-window (1,024) + 8 global layers, vision encoder
- Provider
- Thinking Machines
- Licence
- Apache-2.0
- Availability
- Weights downloadable
- Practical deployment
- Multi-GPU server
- Specified at
- NVFP4 + MXFP8 (mixed)
- System RAM
- 512 GB
- Record updated
- 2026-11-21
Commercial use permitted. No usage restrictions beyond attribution.
Limits and caveats
- Thinking Machines states at least 500 GB of aggregate GPU memory for the NVFP4 checkpoint (for example 2x B300 or 4x H200), and 800 GB for MXFP8.
- The repository has no LICENSE file; Apache-2.0 is stated in the card's metadata and text.
- Benchmarks on the card are vendor-reported.
See also Inkling Small
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run Inkling 2 Small on my GPU?
Other Inkling sizes
Inkling 2 Small — frequently asked questions
How much VRAM does Inkling 2 Small need?
About 394 GB at NVFP4 + MXFP8 (mixed) — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Inkling 2 Small run on an RTX 4090 (24 GB)?
No. Inkling 2 Small needs about 394 GB at NVFP4 + MXFP8 (mixed), more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Inkling 2 Small locally?
Ollama lists Inkling 2 Small only as the cloud tag inkling-2-small:cloud, which runs on Ollama's servers, not your machine. Locally, Thinking Machines documents vLLM and SGLang on a multi-GPU server. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does Inkling come in?
Inkling (NVFP4) (589 GB), Inkling (BF16) (589 GB), Inkling Small (161 GB), Inkling 2 Small (394 GB), Inkling 2 Mini (21 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.