Inkling 2 Mini released by Hugging Face for local AI
Hugging Face has released Inkling 2 Mini, a small multimodal model for local computers with a one million token context.

Key takeaways
- Inkling 2 Mini has 26 billion physical parameters and 3 billion active parameters per token.
- It supports a one million token context and accepts both text and images.
- Hugging Face says it tops 32 of the 48 benchmarks against larger rivals.
- It is available in 4-bit and 8-bit quantised versions for local hardware.
In this article
What Hugging Face announced
Hugging Face has released Inkling 2 Mini, a compact artificial intelligence model designed to run on local hardware. The Hugging Face Blog announced the model on 20 November 2026. It is a multimodal mixture-of-experts system, meaning different parts of the model handle different tasks, and it accepts both text and images.
Hugging Face describes Inkling 2 Mini as a looped transformer model. It has roughly 26 billion physical parameters in total, but only 3 billion are active for each token. The design reuses the middle twelve layers three times, so the model behaves as if it had 48 logical layers while storing just 24 physical ones.
The model supports a context window of one million tokens, meaning it can handle very long inputs, and it can process images as well as text. It uses a mixed-precision format that pairs NVFP4 and MXFP8 expert weights. A DSpark drafter is included to speed up inference, or response generation, at a small memory cost.
How the model works

Inkling 2 Mini is smaller than the Inkling 2 Small model and is intended for local AI setups. It uses 24 decoder layers, with layers six to seventeen repeated in a loop. Its hidden size is 2,304, compared with 4,608 in the Small variant.
At each layer visit, the model routes a token to eight experts chosen from 128 available experts. Every six-layer block mixes five sliding-window layers with a single full-attention layer, and the block works with a local window covering up to 1,024 tokens. The model uses 24 query heads and eight key-value heads for attention.
A research variant named TML-d24-varloop-research can use a different number of passes through the loop. Using one or two passes rather than three speeds up inference, though quality may fall slightly. Hugging Face says it has not evaluated this variant in depth.
Benchmarks and coding performance
Hugging Face says Inkling 2 Mini is leading in its class on the Artificial Analysis Intelligence Index v5. The model tops 32 of the 48 benchmarks when compared with several larger rivals, including Gemma 4 31B and Qwen3.8-27B. The benchmark areas cover coding, agentic search, vision, reasoning and safety.
In coding tests, it comes first in FrontierCode v2.0. It matches Qwen3.8-27B on Terminal-Bench 4.0 but performs worse on AutomationBench-AA v1.0.6. The blog notes that AA Briefcase uses an ELO rating, where higher is better. These results come from internal vibe tests, and Hugging Face says more validation is needed.
How to run it locally
Inkling 2 Mini can be run with transformers, Llama App, MLX and SGLang. Llama App is built on llama.cpp and works on macOS and Windows, providing a local OpenAI-compatible API. The release includes GGUF 4-bit IQ4_NL and 8-bit Q8_0 quantised versions.
SGLang offers day-zero support for production deployment on data centre hardware, with SM103 listed as a prerequisite. Users can adjust reasoning effort from none to max, with high set as the default. The model also supports tool calling, allowing it to request external tools and use their output in its response.
Sources
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.



