Thinking Machines Open-Sources Inkling 2 Mini: 5.4B Active Params Rival 27B-Class Models
Thinking Machines open-sourced the model Inkling 2 Mini in November 2026, and Artificial Analysis subsequently published a series of evaluations. Despite only 5.4B active parameters, the model scored 20 on the Artificial Analysis Intelligence Index, ranking 2nd in intelligence—tying the much larger Qwen3.8 27B (xhigh) and Nanbeige4.3-33B, with roughly 5x and 3x fewer active parameters respectively, placing it in the 20-point Intelligence Index tier.
Confirmed
- Cost and speed: $0.59 and 9.0 minutes per Intelligence Index task, versus $3.5 and longer runtimes for Qwen3.8 27B—about 6x cheaper and 4x faster.
- Compute economics: to reach the same score as Qwen3.8 27B, it outputs about 38% more tokens (about 15.2 per task), but total compute is about 2.8x cheaper, with the efficiency gain driven mainly by the recurrent layers design.
- Artificial Analysis has released full evaluation details for each subcategory (Briefcase, Omniscience, Automation Bench, Terminology, etc.).
Why it matters
- Inkling 2 Mini demonstrates a "small active parameter count + recurrent layers for efficiency" path: matching 27B-class models with roughly 1/5 the active parameters, dramatically cutting inference costs.
- However, the evaluation also notes the trade-off—output is noticeably verbose, requiring more tokens to reach the same score, so real-world usage requires weighing verbosity against cost advantages.
2026-11-21 ~ 2026-11-21 · 7 related posts
- Episode 1: Thinking Machines Launches Inkling 2 Small, Weights to Open-Source Next Week(2026-11-06, 101 posts)
- Episode 2: Inkling 2 Small generates full stock diligence memo for $0.69(2026-11-06, 2 posts)
- Episode 3: Inkling 2 Small tops US open-weight model rankings(2026-11-06, 2 posts)
- Episode 4: Ex-OpenAI Policy Chief Miles Brundage Praises Inkling 2 Small(2026-11-06, 2 posts)
- Episode 5: Inkling 2 Small Pushes Open-Model Frontier in Code and Finance(2026-11-07, 2 posts)
- Episode 6: Leaks claim Thinky's new model matches DeepSeek at a fraction of the cost(2026-11-07, 2 posts)
- Episode 7: Inkling 2 Small Architecture Revealed: 594B Params, 17B Active(2026-11-08, 2 posts)
- Episode 8: Thinking Machines Offers Inkling 2 Small Free on OpenCode(2026-11-08, 4 posts)
- Episode 9: Thinking Machines Releases Inkling 2 Small with Open Weights Coming(2026-11-09, 3 posts)
- Episode 10: Thinking Machines Launches Inkling 2 Small on OpenRouter(2026-11-10, 2 posts)
- Episode 11: Thinking Machines Lab Open-Sources 26B Looped Transformer Inkling 2 Mini(2026-11-21, 51 posts)
- Episode 12: Inkling 2 Mini lands on Ollama, runs locally in 16GB of RAM(2026-11-21, 5 posts)
- Episode 13: Thinking Machines Open-Sources Inkling 2 Mini: 5.4B Active Params Rival 27B-Class Models(2026-11-21, 7 posts)
Primary sources
- Inkling 2 Mini ranks #2 on AA Intelligence Index at 20, but is extremely verbose — ArtificialAnlys ·
- Inkling 2 Mini costs ~6x less and runs ~4x faster per task than Qwen3.8 27B, AA finds — ArtificialAnlys ·
- Inkling 2 Mini Hits 49.3% on ARC-AGI-2 at $0.14 per Task — teortaxesTex ·
- Thinking Machines' open-weights Inkling 2 Mini hits 20 on AA Intelligence Index with just 5.4B active params — ArtificialAnlys · 2026-11-21
- Inkling 2 Mini matches 27B-33B rivals on intelligence with ~5-6x fewer active parameters — ArtificialAnlys · 2026-11-21
- [source] Inkling 2 Mini costs ~6x less and runs ~4x faster per task than Qwen3.8 27B, AA finds — ArtificialAnlys · 2026-11-21
- Looped layers let Inkling 2 Mini use ~2.8x less compute despite 38% more output tokens — ArtificialAnlys · 2026-11-21
- [source] Inkling 2 Mini ranks #2 on AA Intelligence Index at 20, but is extremely verbose — ArtificialAnlys · 2026-11-21
- Full sub-benchmark breakdown for Inkling 2 Mini published by Artificial Analysis — ArtificialAnlys · 2026-11-21
- [source] Inkling 2 Mini Hits 49.3% on ARC-AGI-2 at $0.14 per Task — teortaxesTex · 2026-11-21