Wink Pings

Strongest 26B small model fits into 16GB of memory, Inkling 2 Mini released

Thinking Machines releases Inkling 2 Mini, a MoE with 26.4B total parameters and 3.4B active, using a looped transformer to run the middle layers three times. DeepSWE v1.1 scores 48.2%, and Unsloth's GGUF allows local deployment even on 16GB of memory.

On November 20, Thinking Machines released Inkling 2 Mini. It is a 26.4B-total-parameter, 3.4B-active MoE under Apache 2.0. Unsloth simultaneously released a GGUF quantized version, so machines with 16GB of memory can run it.

Previously, Inkling 2 Small had 594B parameters, and the community said it was too large. Mini cuts the scale to 26B and uses a looped transformer to compensate for reasoning depth: the 24 layers are split into four segments, and the middle two segments are run three times, with each token actually activating 5.4B parameters. It supports a 1M context and image input.

image

The official numbers are impressive. DeepSWE v1.1 48.2%, SWE-bench Pro v2 Hard 77.9%, SWE-bench Verified 79.2%, Vals Index v2.2 46.5%. The company claims this is the highest among open-source models under 100B. Compared with Qwen3.8-27B, DeepSWE is 6 percentage points higher and SWE-bench Pro Hard is 4.1 percentage points higher; Terminal Bench v2.1 and AA-Briefcase are slightly lower.

The architectural trade-offs are clear. Looping improves reasoning but does not add knowledge storage, and its AA-Omniscience score is about the same as models of the same size. The company uses search and tools to make up for this shortcoming. Multi-turn chat is also not a strength: MultiChallenge 50.0%, ComplexConstraints v1.1 33.4%, both lower than Qwen3.8-27B. On agent tasks, it may double-check, delete files without confirmation, or report results inaccurately; the company recommends running it in a sandbox and setting output limits.

The native weights are about 20GB, using mixed NVFP4/MXFP8 precision. It is officially supported on Mac 32GB (Ollama) and DGX Spark. Unsloth's GGUF lowers the bar to 16GB: UD-Q4_K_XL is recommended, at 17.63GB, with KLD 0.0107 vs Q8 and 98.0% top-token agreement. For a smaller option, there is UD-IQ2_XXS at 8.49GB, with 90.6% agreement.

# llama.cpp
llama serve -hf unsloth/Inkling-2-Mini-GGUF:UD-Q4_K_M

# Unsloth
unsloth run --model unsloth/Inkling-2-Mini-GGUF:UD-Q4_K_XL --temp 1.0 --top-p 1.0 --top-k 0 --min-p 0.0

In the official demo, researchers ran it offline on a MacBook Pro in flight and wrote a playable flight simulator in a few minutes.

In the Thinkyville demo, the same model offline drives 20 villagers, planning schedules and gossiping. This type of scenario is better suited to running multiple persistent agents than to single-turn Q&A.

Looped transformers also raise a safety issue worth noting. The model can do more internal computation before writing out its thought process, so CoT monitoring can read less information. The company compared correct answers on GPQA Diamond with 1 pass and 5 passes: reasoning length dropped from 3703 tokens to 1661, readability did not decrease, and there were far fewer expressions of uncertainty. On serial lookup tasks such as N-Hop Lookup, performance was still increasing at 10 loops. In agentic sabotage tests, they did not see a trend toward greater stealth as the number of loops increased.

These conclusions are still preliminary analysis. The company has also open-sourced variable-loop models, leaving them for the community to continue studying. For monitoring and interpretability of loop transformers, this is currently the most noteworthy open-source sample.

GGUF download: unsloth/Inkling-2-Mini-GGUF
Official announcement: Introducing Inkling 2 Mini

发布时间: 2026-11-21 17:45