BestLLMfor Your hardware. Your LLM. Your call.
The Local Copilot Kit APIOpen data Find my LLM
Model fiche

Inkling

By Thinking Machines · United States

Updated 2026-08-28

chat vision audio moe multilingual
Parameters
975B
License
Apache 2.0
Context
1024k
VRAM (Q4)
566 GB
Released
2026-07-14

Overview

Inkling is Thinking Machines' multimodal MoE model — 975B total parameters with 41B active — handling vision, audio, and text across a 1M-token context. Extremely heavy at ~566GB VRAM in Q4.

When to pick this model

  • Multimodal reasoning across text, image, and audio in one model
  • Processing very long documents or transcripts near the 1M-token limit
  • Research and evaluation on multi-GPU server infrastructure
  • Applications needing audio understanding alongside vision and text

VRAM requirements by quantization

VRAM REQUIRED (GB)256512Q4_K_M566 GBQ5_K_M692 GBQ8_01043 GBFP161950 GB
QuantizationVRAM required
Q4_K_M (recommended)566 GB
Q5_K_M692 GB
Q8_01043 GB
FP16 (no quantization)1950 GB

VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.

In practice, Inkling is server-class even at Q4_K_M (566 GB). Stepping up to Q8_0 nearly doubles the footprint to 1043 GB, and unquantized FP16 weights take 1950 GB — plan your GPU around the Q4 or Q5 figure unless you specifically need the higher fidelity.

Without a GPU, Inkling needs roughly 1268 GB of system RAM to run on CPU via llama.cpp or Ollama — workable for background jobs, but far slower than GPU inference. Throughput estimates from our compatibility engine: around 1.5 tokens/sec on entry-level GPUs, on the order of 2.5 tokens/sec on a mid-range card, and up to 5 tokens/sec on high-end hardware — assuming the chosen quantization fully fits in VRAM.

What hardware do you need

The table below matches Inkling to common GPU memory tiers, using the highest-fidelity quantization that fully fits each card class. Spilling layers to system RAM works but costs most of the speed, so size your card to the quantization you actually want to run.

GPU memoryExample cardsBest fit for Inkling
8 GBRTX 5070 Laptop, RTX 5060, RTX 5060 Ti 8GBDoes not fit — needs 566 GB at Q4_K_M
12 GBRTX 5070, RTX 5070 Ti Laptop, RTX 4080 LaptopDoes not fit — needs 566 GB at Q4_K_M
16 GBRTX 5080, RTX 4080 Super, Radeon RX 9070 XTDoes not fit — needs 566 GB at Q4_K_M
24 GBRTX 4090, Radeon RX 7900 XTX, RTX 5090 LaptopDoes not fit — needs 566 GB at Q4_K_M
32 GBRTX 5090Does not fit — needs 566 GB at Q4_K_M

Which GPU should you buy to run Inkling?

To run Inkling locally at Q4, you need ~566 GB of VRAM. The best value for this is a Apple Mac Studio (64+ GB unified memory).

Check Apple Mac Studio price on Amazon →Check Apple Mac Studio price on Newegg →

As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.

Strengths

  • Full multimodal coverage: image, audio, and text
  • Massive context window up to 1M tokens
  • MoE efficiency: only 41B of 975B parameters active per token
  • Apache 2.0 license with no usage restrictions

Limitations

  • Enormous footprint at ~566GB VRAM in Q4 — multi-GPU servers only
  • Slow local throughput (~5 tok/s at Q4)
  • No Ollama tag — install via Hugging Face

Typical workloads

In our catalog grid, Inkling is filed under Multimodal Reasoning, Audio & Vision, Long Documents — the use cases where its size/quality trade-off makes the most sense. Its tags translate to concrete workloads: vision-language work — screenshots, charts, scanned documents; audio understanding; multilingual workloads.

The 1024k-token context window is large enough to hold entire codebases' worth of files or long reports in a single prompt, which is what makes local RAG and document analysis practical. The Apache 2.0 license is permissive, so shipping it inside a commercial product raises no special legal questions.

Architecture & training

Architecture: Multimodal MoE · 975B total parameters, 41B active · 1M-token context window

Training: A multimodal MoE model from Thinking Machines (vision, audio, text). Training details not published.

Verdict

A frontier multimodal MoE model with genuine audio and vision capability, but realistically deployable only on multi-GPU infrastructure.

Quick start

# HuggingFace : thinkingmachines/Inkling

Or use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.

Similar models worth comparing

Frequently asked questions

How much VRAM does Inkling need?

At the recommended Q4_K_M quantization, Inkling needs about 566 GB of VRAM. Q8_0 takes 1043 GB, and unquantized FP16 weights take 1950 GB.

Can Inkling run without a GPU?

Yes — with roughly 1268 GB of system RAM it runs CPU-only through llama.cpp or Ollama. Expect a fraction of GPU speed, which is fine for background or batch jobs but slow for interactive chat.

What context window does Inkling support?

Inkling supports a 1024k-token context window (1,048,576 tokens).

Can I use Inkling commercially?

Yes. Inkling is released under Apache 2.0, a permissive open-source license that allows commercial use, modification and redistribution.

How fast is Inkling on consumer hardware?

Our compatibility engine estimates on the order of 2.5 tokens/sec on a mid-range GPU and up to 5 tokens/sec on high-end cards, assuming the quantization fully fits in VRAM.

Which quantization of Inkling should I download first?

Start with Q4_K_M (566 GB) — the standard size/quality sweet spot. Step up to Q5_K_M or Q8_0 only if you have VRAM headroom. It does not fit a single 24 GB consumer card — plan for multi-GPU or server hardware.

Tools

Is Inkling the right pick for you?

Compute self-hosted ROI → Back to catalog