Skip to main content

Thinking Machines presents Inkling 2 Mini for local inference

21 NOVEMBER 2026·2 MIN READ·4 SOURCES

Thinking Machines describes Inkling 2 Mini as a fast model small enough to run on a laptop. It accepts text and images, generates text, and supports a context window of up to 1 million tokens.

Key takeaways · 3

  • 01

    The model has a 1-million-token context window and handles text and image inputs, with text output.

  • 02

    Thinking Machines lists Apple silicon devices with 32GB or more memory; the model checkpoint is about 20GB.

  • 03

    Tinker offers API access, and the release supports Transformers, Llama App and MLX; SGLang supports datacenter inference.

Model design and capabilities

Inkling 2 Mini is a multimodal model that accepts text and images and generates text, with a context window of up to 1 million tokens.[2] It has 26.4 billion total parameters and 3.4 billion active parameters.[1] Its architecture has 24 layers, with the middle 12 running three times per token.[2] Thinking Machines says this looping lets the model apply more reasoning and reach the right answer with fewer tokens.[1] The weights use NVFP4 and MXFP8 formats, and the checkpoint is about 20GB.[2]

Access and hardware considerations

Developers can access Inkling 2 Mini through Tinker, which provides API access; Thinking Machines says the model is available there for inference and customization.[2][1] The release supports Transformers, Llama App (llama.cpp) and MLX, while SGLang provides day-zero inference support for production deployments on datacenter hardware.[3] Thinking Machines describes the model as small enough to run on a laptop and lists Apple silicon devices with 32GB or more memory.[1][2] The release also includes a DSpark drafter for additional speed-up at a small memory cost.[3]

Reported benchmark results

Thinking Machines says Inkling 2 Mini scored 48.2% on DeepSWE v1.1, leading open-weight models below 100 billion total parameters on that benchmark.[1] It reports a 46.5% score on Vals Index v2.2.[1] Hugging Face says the model led in 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL.[3] Results were not uniform: Thinking Machines reports scores of 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, trailing Qwen3.8-27B on both.[1]

Cost estimates and caveats

Artificial Analysis estimated an Intelligence Index task cost of $0.59 and completion time of 9 minutes for Inkling 2 Mini, compared with $3.53 and 38.2 minutes for Qwen3.8 27B.[4] Its cost estimate used Thinking Machines’ first-party prices of $0.15 per million input tokens, $0.50 per million output tokens and $0.015 per million cache-hit tokens.[4] Thinking Machines says its safety evaluations found the model materially below frontier capabilities on loss-of-control assessments.[2] These reported results and estimates are useful inputs for evaluation, not a substitute for testing the model against a team’s own workloads and hardware.

The combination of local tooling, API access and a long context window gives technical teams several deployment paths to evaluate. Benchmark claims, memory needs and task-specific performance should be checked against the workloads and hardware a team actually plans to use.

Why it matters
For teams

AI training for your team

kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.

How this developed

  1. 21 November 2026

    Thinking Machines presents Inkling 2 Mini for local inference

Sources

AI fluency, one session a day, built for your work.