This observation covers AI technology developments formally released or newly verified during the approximately 24 hours before 7:49 am on 22 November 2026 in Sydney, Australia. One item is included.
Inkling 2 Mini open-weight release and independent testing
Thinking Machines Lab released Inkling 2 Mini on 20 November with downloadable weights under the Apache 2.0 licence. It is a mixture-of-experts model that accepts text and image inputs and produces text. The model has 26.4 billion total parameters and activates 3.4 billion for each token; because its middle layers run three times per token, the developer counts 5.4 billion effective weight activations. It supports a context window of up to one million tokens. Most of the weights use NVFP4 precision and the final four layers use MXFP8, with the developer reporting a memory footprint of about 20 GB. Compared with the larger Inkling 2 Small that was already available, this delivery adds a smaller model intended to run locally on personal workstations and laptops, together with a research variant whose number of loops can be changed. The weights are available through Hugging Face, Tinker provides inference and fine-tuning, and launch support includes runtimes such as SGLang, vLLM and Ollama.
The model card identifies long-running agent tasks, tool use, coding and retrieval as primary uses, while also recording specific limits. Knowledge and factual accuracy are weaker than Inkling 2 Small, performance can degrade in long conversations, code debugging can enter repetitive reasoning loops, and tool use can delete or overwrite files without confirmation. Output quality can also decline outside English. The model card recommends sandboxing agent workloads and requiring confirmation before irreversible actions. The developer reports 48.2 per cent on DeepSWE v1.1; this is a developer-run evaluation and is not treated as a broadly reproduced result.
On 21 November, Plotly published an independent test using a consistent data-analytics agent harness. Inkling 2 Mini answered 36 of 43 questions correctly. The complete run cost US$4.09, took a median of 69 seconds per question and used 179 million tokens. One hidden-pattern task triggered 585 queries over about 48 minutes and accounted for 46 per cent of all tool calls in the test, yet the model still missed the answer. Plotly notes that it ran each model once and that repeat runs of other models have moved by as many as three points, so the score should not be read as a stable ranking. The test independently confirms that the model can complete multi-step tool use in an existing hosted environment and records a case in which looped reasoning increased cost and latency. Public independent reproduction across different local hardware, languages and a broader range of agent tasks is not yet available.
Sources
Thinking Machines Lab, Introducing Inkling 2 Mini, 20 November 2026, https://thinkingmachines.ai/news/inkling-2-mini/
Thinking Machines Lab, Inkling 2 Mini Model Card, 20 November 2026, https://thinkingmachines.ai/model-card/inkling-2-mini/
Plotly, Inkling 2 Mini Benchmark Results for Data Analysis, 21 November 2026, https://plotly.com/blog/inkling-2-mini-plotly-data-analytics-bench/
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.