Evidence: benchmark results vendor-reported
Inkling 2 Mini — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
Thinking Machines' first looped model: 24 layers whose middle 12 run three times per token, so 26.4B parameters (3.4B active, about 5.4B effective with the loops) do the work of a deeper network. It takes text and images, has a 1M-token context, and ships as one ~20 GB checkpoint in mixed NVFP4/MXFP8 that runs on a DGX Spark or a Mac with 32 GB or more. Ollama carries it as inkling-2-mini:26b-a5b, and llama.cpp loads the ggml-org GGUF files.
Inkling 2 Mini needs about 21 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.
Get Inkling 2 Mini running
The cheapest catalogued GPU that runs Inkling 2 Mini is the AMD Radeon RX 7900 XTX (24 GB).
or compare on Vast.ai from $0.35/hr (typical low · varies)
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 2 Mini
Install Ollama, then run:
ollama run inkling-2-mini:26b-a5bSpecifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 26.4 Billion (3.4B active, ~5.4B effective)
- Context window
- 1,048,576
- Architecture
- Looped Mixture-of-Experts (128 experts, 8 active): 24 layers, middle 12 run 3x per token; vision encoder
- Provider
- Thinking Machines
- Licence
- Apache-2.0
- Availability
- Weights downloadable
- Practical deployment
- Single consumer GPU
- Specified at
- NVFP4 + MXFP8 (mixed)
- System RAM
- 32 GB
- Record updated
- 2026-11-21
Commercial use permitted. No usage restrictions beyond attribution.
Limits and caveats
- Looped layers run three times per token and each pass keeps its own cache, which this site's memory and speed model cannot yet express: memory shown is the published file size, with no speed estimate.
- Thinking Machines warns it can get stuck in long reasoning loops, especially when debugging code; set limits on output tokens and agent turns.
- Released 20 November 2026; benchmarks on the card are vendor-reported.
See also Inkling 2 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 Mini on my GPU?
Other Inkling sizes
Inkling 2 Mini — frequently asked questions
How much VRAM does Inkling 2 Mini need?
About 21 GB at NVFP4 + MXFP8 (mixed) — quantized weights plus framework overhead, before any KV cache. We publish no context-length figure or speed for it: its middle 12 layers run three times per token and each pass keeps its own cache, so the bytes read per token are not the weights once, which our memory and speed model cannot yet express. Apple Silicon counts unified memory toward the same figure.
Does Inkling 2 Mini run on an RTX 4090 (24 GB)?
Its weights and overhead come to about 21 GB at NVFP4 + MXFP8 (mixed), but we do not publish a fit verdict or a speed for it: its middle 12 layers run three times per token and each pass keeps its own cache, so the bytes read per token are not the weights once, which our memory and speed model cannot yet express.
How do I run Inkling 2 Mini locally?
Install Ollama and run `ollama run inkling-2-mini:26b-a5b`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the 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.