Inkling is Thinking Machines' single 975B MoE release at 1M context, Apache-2.0. A frontier-scale model — cloud or datacenter territory, listed here for completeness.
Inkling is a single-model family from Thinking Machines, released in 2026 as a 975B-parameter sparse MoE under the permissive Apache-2.0 license. It has one member and one purpose: to sit at the extreme end of what open-weight language models can be. With 975B total parameters, a 41B active footprint per token, and a 1M context window, it is the second-largest model in our catalog — behind only the trillion-parameter Kimi K2, and with a more permissive license.
The honest headline: this is a datacenter-only model. No single consumer, workstation, or even unified-memory workstation GPU holds Inkling at any quant we track. This page tells you exactly how much VRAM is required, so you know whether your cluster clears the bar.
Size ladder
- Inkling (975B MoE, 41B active, 1024K ctx, Apache-2.0) — Q4 weights alone are ≈ 536.3 GB. With KV cache at 8K context that is roughly 537.3 GB total → you need roughly 6×96GB GPUs (or equivalent datacenter cards) just to hold a Q4 quant at 8K context. See detailed VRAM matrix
What makes Inkling different
Inkling is here as a reference point, not a shopping recommendation — but the specs are worth understanding:
- 41B active, 975B total. The MoE ratio is about 24:1, meaning per-token compute cost tracks 41B while the model stores nearly a trillion parameters. This is a meaningful efficiency gain — once you actually fit the model somewhere.
- 1M context window. Tied with the longest we track. At the full 1M context, KV cache alone would be roughly 1,025 GB — effectively doubling the 8K baseline VRAM. Running the 1M window at Q4 means north of 1.5 TB of aggregate VRAM.
- Apache-2.0 license. This is the real differentiator: a near-trillion-parameter model under one of the most permissive open-source licenses. Commercial use, fine-tuning, redistribution — all on the table, legally. The bottleneck is purely hardware.
What hardware runs Inkling
No single GPU we track can hold Inkling — not even a 192GB unified-memory chip. Realistic tiers:
- Q2 quant (≈274 GB weights): 3×96GB cards, or a pair of H200 (141GB) cards with offload. Usable but expect quality loss from aggressive quantization.
- Q4 quant, 8K context (≈537 GB): 6×96GB-class cards, or 4× H200. This is the minimum for full-quality Q4 at modest context.
- Q4 quant, full 1M context (≈1.56 TB): 16×96GB-class cards. Long-context KV cache is the driver; at this tier you are in HPC-cluster territory.
For the overwhelming majority of builders, the practical takeaway is: Inkling exists, it is MIT-style permissive, and it shows where the frontier is. If your cluster can run it comfortably, you already know. Everyone else is better served by a model in the 70B–200B range that fits one or two consumer cards — start with our 12GB, 16GB, or 24GB best-model lists.
Links
1 models in the Inkling family, grouped by series. Q4 (GB) is weights-only; total VRAM adds KV cache and overhead — the fit check links compute it for a representative retail GPU at 8K context.