Tinker Results Don't Prove Inkling is the Best

hero88645 · reddit · 2026-07-19

The author reviews public materials related to Thinking Machines' Tinker, questioning whether the best public results actually prove that Inkling is the optimal foundation model.

Public Benchmarks

On AIME 2026, Inkling scores 97.1%, but higher-scoring models exist in the same table, such as GLM 5.2 at 99.2%, and Fable 5 and GPT-5.6 Sol at 99.9%.
On HLE text-only, Inkling scores 29.7%, beating Nemotron 3 Ultra but trailing GLM 5.2, DeepSeek V4 Pro, and two Kimi models.
On SWEBench Pro and Terminal Bench 2.1, the trend continues: Inkling beats some models but falls behind stronger open/closed-source alternatives.
Its only standout performance is on IFBench, where it leads with 79.8%, indicating strong instruction-following capabilities.

Core Controversy

The author points out that people often conflate two things:
Tinker can fine-tune open-source models into highly capable specialized models.
Inkling itself is the best foundation model to use.
Current public evidence only supports the first point, not the second.

The Bridgewater Case

The mentioned Bridgewater/Tinker case study used Qwen3-235B as the base, not Inkling.
This case achieved an 84.7% accuracy across six financial document filtering tasks, with per-task inference costs roughly 13.8x lower than tested frontier models.
Inkling did not exist when this case study was published, so it cannot be used to prove Inkling's superiority.

Conclusion

The author concludes that current materials only prove Tinker's fine-tuning/specialization pipeline is effective, but there is insufficient evidence to crown Inkling as the go-to foundation model.

Original post →

Tinker Results Don't Prove Inkling is the Best

More from Models