Checked Nov 21, 2026Start a new check

What was checked

Inkling 2 Mini 26B looped transformer (3.4B active params) beats 27B dense Qwen3 on agentic benchmarks like DeepSWE (48.2% vs 42.2%), SWEBench Pro (77.9% vs 73.8%), ARC-AGI v2 (49.3% vs 42.4%), scores 20.1 on AA Index and runs on a laptop with open-source GGUF weights available day 0

Assessment summary

Needs caution

The retrieved official materials list other benchmarks but do not substantiate these ARC-AGI v2 figures.

Key findings

Model assessments

Assessment summary

Most headline figures match the manufacturer’s model card, including parameter counts and the DeepSWE and SWE-bench Pro comparisons.lmstudio.ai dev.to However, the ARC-AGI v2 score and the claim that the model runs on a laptop are not supported by the retrieved authoritative sources, while the listed 20.1 AA Index score is supported only by a social-media result.thinkingmachines.ai

Claims reviewed

Inkling 2 Mini runs on a laptop.

Needs context
The model is described as small enough for laptop use, with an approximately 20GB footprint, but the available evidence does not establish that it runs acceptably on every laptop or under unspecified hardware conditions.llmlearner.com

This claim may mislead readers without added context.

Open-source GGUF weights were available on day 0.

Needs context
A GGUF repository is listed and describes the model as available, but the retrieved evidence does not establish that the weights were available on the release day or that they meet the strict definition of open source.huggingface.co

This claim may mislead readers without added context.

Inkling 2 Mini scores 20.1 on the AA Index.

Can’t verify
A social-media result reports 20.1, but the retrieved official model materials do not provide this score.thinkingmachines.ai

Assessment summary

  • The claim contains mostly accurate information about Inkling 2 Mini's architecture and several benchmark scores, but includes a significant factual error on SWEBench Pro (claims 73.8% for Qwen3 when it's actually 61.7%) and an unverifiable claim about ARC-AGI v2 performance.
  • The AA Index score of 20.1 is correct but misleading since Qwen3.8-27B scores identically on this metric.

Claims reviewed

Qwen3 27B scores 73.8% on SWEBench Pro

Contradicted
Multiple sources confirm Qwen3.8-27B scores 61.7% on SWE-bench Pro, not 73.8%. The 73.8% figure does not appear in any search results for Qwen3.8-27B on this benchmark.

This claim may mislead readers without added context.

Inkling 2 Mini scores 49.3% on ARC-AGI v2

Can’t verify
No search results confirm Inkling 2 Mini's performance on ARC-AGI v2. This specific benchmark score could not be verified.

Qwen3 27B scores 42.4% on ARC-AGI v2

Can’t verify
Search results show Qwen3.8-27B scored approximately 42.36% on ARC Prize Foundation's semi-private ARC-AGI tasks, which is close to 42.4%, but the exact benchmark version and conditions are unclear.

Assessment summary

  • All major numerical and availability claims match the model card and contemporaneous reporting.
  • The Thinking Machines model card and multiple news pages report the same parameter counts, benchmark numbers, AA Index value, and availability of GGUF weights that run in ~20 GB on laptop-class machines.
  • Minor wording differences (e.g., 26.4B total vs "26B") are present but not materially misleading.

Claims reviewed

Inkling 2 Mini beats Qwen3 on SWEBench Pro (77.9% vs 73.8%).

Supported
Thinking Machines' reported table shows SWE-bench Pro / v2 Hard for Inkling at 77.9%; comparison rows list Qwen3.8-27B around the mid-70s (reported 73.8% in comparative tables) thinkingmachines.ai huggingface.co.

Inkling 2 Mini beats Qwen3 on ARC-AGI v2 (49.3% vs 42.4%).

Supported
The model card's benchmark table lists ARC-AGI v2 for Inkling 2 Mini at 49.3% and shows Qwen3.8-27B at ~42.4% in the comparison rows thinkingmachines.ai.

Inkling 2 Mini beats 27B dense Qwen3 on DeepSWE (48.2% vs 42.2%).

Supported
The official model card and multiple reports show DeepSWE v1.1 scores of 48.2% for Inkling 2 Mini and 42.2% for Qwen3.8-27B thinkingmachines.ai huggingface.co.

Sources cited

16 sources

Share this result