Skip to content
English

Switch theme

// Thinking Machines Inkling · Vision-language

Inkling: VRAM & local deployment

975B total · 41B active (MoE) parameters. About 651 GB at 4-bit and 2.15 TB at BF16.

released 2026-07-14licence: Commercial OKmemory figures are estimates

01Specs

PublisherThinking Machines Lab · thinkingmachines
Family / seriesThinking Machines Inkling · Inkling
TypeVision-language (image-text-to-text)
Parameters975B total · 41B active (MoE) Source: official model card“975B total, 41B active”
Context length—
Architecture66 layers · GQA · MoE
Released weightsBF16 · 1.90 TBdownload
LicenceApache-2.0 · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed.
Released2026-07-14 (Hugging Face repo created)

02How much VRAM it needs (estimate)

Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.

PrecisionWeightsSmallest GPUMac
BF16 / FP16 (original)1.95 TB 2 × 8 × B300 nodes (4,608GB)—
FP8 / INT8 / Q8_01.04 TB 8 × B200 (1,440GB)—
4-bit (Q4_K_M / AWQ / GPTQ-Int4)591 GB 8 × H200 (1,128GB)—

KV cache not estimated: this model uses a new compressed / sparse attention design that our formula does not cover. Totals above are weights + overhead only.

Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →

03How to run it

Only runtimes the publisher or the runtime’s official library names for this model.

Which inference engine to choose →

04Official repositories & quantised builds

2 official repo(s) from thinkingmachines.

05Other Thinking Machines Inkling models

All 2 →
ModelParams4-bit totalReleased
Inkling-Small276B / 12B185 GB2026-07-27

06Sources

Data checked 2026-07-23. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

Data updated: Self-host open models(UTC+8)