AI Aulis

AI NEWS IN ENGLISH · 21 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

← Issue 21 November 2026Paper

Inkling 2 Mini matches Qwen3.8 27B on Artificial Analysis’s index

AI-generated contentWritten by AI

huggingface.co, vllm.ai, github.com, goodfire.com, artificialanalysis.ai, x.com, superpowerdaily.com · 8 sources · Sources published: 20–21 November 2026

Thinking Machines’ open-weight model Inkling 2 Mini scored 20 points on the Artificial Analysis Intelligence Index, the same as Qwen3.8 27B. The model runs its middle layers three times, so its active parameter count depends on how it is counted.

Thinking Machines released Inkling 2 Mini on 20 November, and its weights are available to download from Hugging Face. Artificial Analysis said the model has been released under an Apache 2.0 licence. It accepts text and images and has a context window of 1 million tokens. Inkling 2 Mini is a mixture-of-experts (MoE) model: it uses only some of its 26 billion parameters for each token.

What sets the model apart is a loop. The model has 24 layers, Hugging Face’s release blog says. The first six and last six layers run once, while the middle 12 run three times with the same weights. On the second and third passes, the layers receive messages from the previous pass. The repetition adds compute, but the model does not need to store new weights.

As a result, the number of active parameters depends on how they are counted. Hugging Face gives about 3 billion active parameters and the vLLM blog 3.5 billion. Artificial Analysis counts all three passes and arrives at 5.4 billion. It says the model uses 3.1 billion distinct weights per token. Artificial Analysis bases its comparison on the larger figure: by that measure, Inkling 2 Mini matches Qwen3.8 27B with about a fifth of Qwen’s active parameters.

The table shows the differences Artificial Analysis measured on the same index tasks.

Metric Inkling 2 Mini Qwen3.8 27B (xhigh)
Intelligence Index score 20 20
Output tokens per task c. 152,000 c. 110,000
Cost per task $0.59 $3.53
Time per task 9.0 min 38.2 min

Source: Artificial Analysis. Inkling’s cost is calculated using Thinking Machines’ own API prices.

The lower price does not tell you what it costs to run the model on your own machine. Inkling 2 Mini also produces about 38 per cent more output tokens than Qwen3.8 27B. Even so, it uses about 2.8 times less compute on Artificial Analysis’s compute measure, which multiplies active parameters by the number of tokens. Thinking Machines itself says the model falls behind Qwen3.8 27B in multi-turn conversations and in instructions with many constraints, the news site Superpower Daily reported.

Artificial Analysis said the model is small enough to run on a MacBook or a DGX Spark. Supported tools include transformers, llama.cpp, MLX, vLLM and SGLang. Support arrived in llama.cpp version b12600, and vLLM has released a dedicated Docker image, vllm/vllm-openai:inkling-2-mini. Hugging Face offers 4-bit and 8-bit versions for llama.cpp. A separate research version lets users cut the passes to one or two, which makes the model faster, though its answers may get slightly worse. Hugging Face has not yet evaluated this version thoroughly.

Previous coverage

Report an error in this story