FULL STORY

Inkling 2 Small: Small Model Shakes Up Open-Weight Rankings

Thinking Machines released the open-weight Inkling 2 Small on Nov 6, topping Design Arena's open-weight leaderboard and matching Kimi-K3 on agentic tasks, sparking a wave of same-day benchmarks and tests.

2026-11-06 ~ 2026-11-06 · 5 episodes · 52 posts

Episode 1 · Thinking Machines Releases Inkling 2 Small Open-Weight Model (2026-11-06, 42 posts)

Thinking Machines, the company founded by Mira Murati, released its second-generation model Inkling 2 Small on November 6 and announced progress on its "model factory." Following Inkling Small, the release focuses on long-horizon autonomous tasks and post-trainability.

Confirmed

  • Architecture and capabilities: a MoE architecture with 594B total parameters and 17B activated, supporting image input and variable effort; built for long-horizon autonomous tasks, able to visually inspect its work and iterate, and adaptable to different agent frameworks and environments
  • Benchmarks: scored 31 on the Artificial Analysis Intelligence Index v5.0.1, up 18 points from the previous-generation Inkling Small
  • Availability: fine-tuning and inference are now open on the Tinker platform—users can fine-tune it themselves or let the model autonomously orchestrate the entire fine-tuning workflow; full weights will be open-sourced soon
  • Training efficiency: per engineering lead cHHillee, cited in a repost, pretraining took less than 3 days; developer simonguozirui reported that 75 steps of fine-tuning on Tinker matched the previous generation's final score, with fine-tuning efficiency far ahead across multiple post-training scenarios using the same data and algorithms
  • Positioning: if the weights are open-sourced as planned, it would become the strongest open-weight model outside China

Why it matters

  • "Post-trainability" is the core selling point: the release emphasizes developer efficiency in fine-tuning over conventional benchmark scores, which directly affects the practical value of open-weight models
  • Reaching this performance with less than 3 days of pretraining (if accurate) showcases the rapid iteration of Thinking Machines' "model factory," suggesting notably faster release cadence going forward
  • If it truly becomes the strongest open-weight model outside China after release, it would reshape the current open-source model landscape with direct impact on developers who rely on open weights

Not yet confirmed

  • Pricing, detailed training specifics, and a full technical report have not been disclosed
  • The "strongest open-weight model" claim rests on a single Artificial Analysis index and requires more third-party evaluation once the weights are actually released

22 more related posts →

Episode 2 · Inkling 2 Small tops Design Arena open-weights ranking (2026-11-06, 3 posts)

Thinking Machines' Inkling 2 Small ranked 4th overall on Design Arena with 1376 Elo, claiming the top spot among open-weight models while its output pricing is about 15 times cheaper than Kimi K3.

Episode 3 · Inkling 2 Small generates full stock diligence memo for $0.69 (2026-11-06, 2 posts)

Inkling 2 Small, a 594B-parameter MoE model from Thinking Machines, generated a complete stock diligence memo for just $0.69 and caught a $1.2 billion error in thoughtfullab's DiligenceBench financial evaluation.

Episode 4 · Thinking Machines Releases Inkling 2 Small as Strongest Open-Weight Model Outside China (2026-11-06, 3 posts)

Thinking Machines released Inkling 2 Small, and Artificial Analysis published full benchmark results under Intelligence Index v5.0.1 with comparison pages, suggesting it may be the strongest open-weight model outside China.

Episode 5 · Thinking Machines' Inkling 2 Small Matches Kimi-K3 on SWE Atlas (2026-11-06, 2 posts)

Thinking Machines released Inkling 2 Small, which leverages well-designed rubrics to rival larger models on agentic tasks, scoring 74.2 and ranking second on the updated SWE-Atlas v1.1 QnA leaderboard, just behind Opus 5.5 and matching Kimi-K3.