Thinking Machines' Open-Source Inkling 2 Mini Matches Frontier Models at a Fraction of Cost
Thinking Machines Lab (TML) released the open-source small model Inkling 2 Mini in November 2026, with 26.4B total parameters and 3.4B activated (effectively around 5.4B). It runs locally on machines with roughly 16–20GB of VRAM/memory or 32GB of RAM. Results from multiple third-party evaluation organizations show that, with a tiny parameter count and cost, the model matches or even outperforms flagship models with an order of magnitude more parameters on several benchmarks, making it the center of this round of discussion.
Confirmed
- Vals AI CyberBench v1.1 (November 20 leaderboard): Inkling 2 Mini tied for 9th–10th place with Claude Opus 5.5 (CVP) at 74.58% accuracy, with a per-test cost of $0.93—about a quarter of Opus's $4.17 (m9/m10/m11/m15).
- Artificial Analysis Intelligence Index score of 20: on par with Qwen3.8 27B (xhigh) and Nanbeige4.3-33B while activating roughly 5x fewer parameters (and more in the latter case); at $0.59 per task and 9.0 minutes, it is about 6x cheaper and 4x faster than Qwen3.8 27B (m1/m3/m4/m13). AA also noted it is "extremely verbose," producing about 38% more output tokens for the same score (15.2K per task), yet total compute is still 2.8x lower, with the efficiency credited to its recurrent-layer design (m5/m13).
- ARC Prize's ARC-AGI(Verified) results: 49.3% on ARC-AGI-2 at $0.14/task; 84%+ on ARC-AGI-1 (m7).
- Energy efficiency: JonSaadFalcon benchmarked Intelligence-per-Watt on a single NVIDIA DGX Spark, reaching 82.3% accuracy at low effort with MXFP8 precision while consuming only 1.17 kJ; Stanford's Hazy Research measured "intelligence per joule" at 0.703, beating both Qwen and Gemma (m16/m17).
- Real-world usage: blogger HowDevelop ran 6 practical capability tests with Inkling-2-Small (hosted on Nebius)—vision, document processing, 3D generation, scoped coding, web research, and knowledge work—finding 5/6 usable, and 3.7–4.9x cheaper than Sonson 5.5 on structured tasks (m12).
- Reddit users report it approaching or even surpassing Gemini 2.5 Pro on multiple benchmarks (m8).
Unconfirmed
- The "17x compute efficiency gain over 128 days" and "outperforming a 975B-parameter, 41BAB model" relayed by teortaxesTex are geometric-mean figures provided by TML; the poster themselves noted their understanding of the data may be inaccurate, and the exact methodology awaits verification (m2/m14).
- The claim in m8 that it "matches Gemini 2.5 Pro on multiple benchmarks" comes from Reddit users and has not been directly corroborated by any evaluation leaderboard.
Why it matters
Inkling 2 Mini demonstrates a route opposite to "stacking parameters": using architectural efficiency (such as recurrent layers) to approach flagship-level intelligence on consumer hardware, while cutting per-task costs to a fraction of the original and minimizing energy consumption. If the data holds up under further independent verification, it could reshape developers' expectations of small models' cost-effectiveness in replacing frontier models for coding, cybersecurity, and automation pipelines—and push "intelligence per joule" as a new model evaluation metric.
2026-11-20 ~ 2026-11-21 · 20 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' Open-Source Inkling 2 Mini Matches Frontier Models at a Fraction of Cost(2026-11-20, 20 posts)
- Episode 12: Thinking Machines Releases Inkling 2 Mini; Unsloth Ships GGUF for Local Use(2026-11-20, 4 posts)
- Episode 13: Thinking Machines Releases 26B Looped Transformer Inkling 2 Mini, Runs on Laptops(2026-11-21, 89 posts)
- Episode 14: Inkling-2 Mini Hits 210 tok/s Locally on Apple Silicon, 11x Cheaper than Sonnet(2026-11-21, 4 posts)
- Episode 15: Inkling 2 Mini local tests: fast and strong at coding, but deployment remains rough(2026-11-21, 6 posts)
- Episode 16: Unsloth Guide: Run Inkling-2-Mini Locally with Just 9GB RAM(2026-11-21, 2 posts)
Primary sources
- Inkling-2-Small costs 3.7-4.9x less than Claude Sonnet 5.5 on structured pipeline tasks — HowDevelop · 2026-11-20
- Inkling 2 Mini ties Claude Opus 5.5 on CyberBench at less than a quarter of the cost — LiTianleli · 2026-11-21
- [source] 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
- 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
- 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
- 26.4B Inkling 2 Mini matches 975B model in 128 days, ~17x compute efficiency gain — teortaxesTex · 2026-11-21
- 26.4B model claimed to beat 975B: 17x compute efficiency gain in 128 days — teortaxesTex · 2026-11-21
- Stanford's intelligence-per-joule test: Inkling 2 Mini scores 0.703, far ahead of Qwen and Gemma — simonguozirui · 2026-11-21
- Reddit claims open-weight Inkling 2 Mini nearly matches Gemini 2.5 Pro on benchmarks — ProxyLumina · 2026-11-21
- [source] Inkling 2 Mini Hits 49.3% on ARC-AGI-2 at $0.14 per Task — teortaxesTex · 2026-11-21
- Thinking Machines' Inkling 2 Mini matches Claude Opus 5.5 at 74.58% on CyberBench — LiTianleli · 2026-11-21
- [source] Inkling 2 Mini ties Claude Opus 5.5 on CyberBench at quarter of the cost — iGermanProd · 2026-11-21
- Inkling 2 Mini tops intelligence-per-joule test: 82.3% accuracy at just 1.17 kJ per query — simonguozirui · 2026-11-21
- Thinking Machines' new model claimed 5x faster than Qwen3.8, 15% more accurate — LiTianleli · 2026-11-21
- Inkling 2 mini: 26.4B MoE with 3.4B active matches Qwen3 27B under Apache 2.0 — LiTianleli · 2026-11-21
- Inkling-2-Mini earns early praise: outperforms Qwen3 and runs much faster — LiTianleli · 2026-11-21
1 near-duplicate retellings: Aubrey_D_Graham