Inkling (NVFP4) — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
The 4-bit deployment build. NVFP4 quantization brings the 975B / 41B-active MoE down to ~600 GB — a single high-memory GPU server rather than a multi-node cluster — while keeping the full 1M-token context and native text / image / audio input. Includes MTP speculative-decoding layers for faster generation and runs under vLLM, SGLang, Transformers 5.14+ and llama.cpp. Blackwell GPUs are recommended for native FP4. Apache 2.0. Thinking Machines reports AIME 2026 97.1%, GPQA Diamond 87.2%, SWE-bench Verified 77.6% and VoiceBench 91.4%. Specs from launch coverage — verify on the Hugging Face model card.
Inkling (NVFP4) needs about 589 GB of VRAM at NVFP4 (4-bit) — 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). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 321.3 GB | — | Too big | |
| Q3_K_M 3.41 bpw | 416.4 GB | — | Too big | |
| Q4_K_M 4.83 bpw | 589.5 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 691.8 GB | — | Too big | |
| Q6_K 6.56 bpw | 800.3 GB | — | Too big | |
| Q8_0 8.50 bpw | 1036.7 GB | — | Too big | |
| F16 16.00 bpw | 1950.8 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 (NVFP4) VRAM calculator →
Get Inkling (NVFP4) running
Inkling (NVFP4) is not practical on a single consumer GPU: it needs about 589.5 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:
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Inkling (NVFP4)
No Ollama library entry. Thinking Machines' documented serving routes are SGLang and vLLM.
Specifications
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong.
- Parameters
- 975 Billion (41B active)
- Context window
- 1,000,000
- Architecture
- Multimodal Mixture-of-Experts — 256 experts, top-6 routed + 2 shared; NVFP4 4-bit
- Provider
- Thinking Machines
- Licence
- Apache 2.0
- Practical deployment
- Private cloud / datacenter cluster
- Specified at
- NVFP4 (4-bit)
- System RAM
- 1024 GB
- Record updated
- 2026-07-16
- Runs with
- Ollama: not supported
Commercial use permitted. No usage restrictions beyond attribution.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
| Benchmark | Score | Provenance |
|---|---|---|
| AIME 2026 | 97.1 % | vendor-claimed · Thinking Machines (launch) |
| GPQA Diamond | 87.2 % | vendor-claimed · Thinking Machines (launch) |
| HLE (with tools) | 46 % | vendor-claimed · Thinking Machines (launch) |
| SWE-bench Verified | 77.6 % | vendor-claimed · Thinking Machines (launch) |
| MMMU Pro (Standard 10) | 73.3 % | vendor-claimed · Thinking Machines (launch) |
| VoiceBench | 91.4 % | vendor-claimed · Thinking Machines (launch) |
| Global-MMLU-Lite | 88.7 % | vendor-claimed · Thinking Machines (launch) |
Can I run Inkling (NVFP4) on my GPU?
Other Inkling sizes
Inkling (NVFP4) — frequently asked questions
How much VRAM does Inkling (NVFP4) need?
About 589 GB at NVFP4 (4-bit) — 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 (NVFP4) run on an RTX 4090 (24 GB)?
No. Inkling (NVFP4) needs about 589 GB at NVFP4 (4-bit), 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 (NVFP4) locally?
No Ollama library entry. Thinking Machines' documented serving routes are SGLang and vLLM. 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). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.