Inkling-Small只用12B激活参数就逼近DeepSeek V4 Flash，开放模型里排21名，少算力也能打，关注后续人投排名。
Inkling-Small在Text Arena获得约第88名，AutoEval得分1431，开放模型中约第21。它仅用12B激活参数（总276B），比肩Minimax M2.7（1417分，10B激活）、Nemotron 3 Ultra（1426分，55B激活）和DeepSeek V4 Flash（1436分，13B激活）。该分数为AutoEval初步结果，由奖励模型代替人工投票产生。
Inkling-Small by @thinkymachines landed in Text Arena with an overall rank of ~#88 (1431 pts, AutoEv...
Inkling-Small by @thinkymachines landed in Text Arena with an overall rank of ~ #88 (1431 pts, AutoEval) ~ #21 among open - but this is with only 12B active params (276B total). This is neck-and-neck with open models running way more compute: - Minimax M2.7: 1417 (10B active / 230B total) - Nemotron 3 Ultra 1426 (55B active / 550B total) - MiMo V2.5: 1434 (15B active/309B total) - Deepseek V4 Flash: 1436 (13B active/ 284B total) - Qwen3.5 397b-A17b: 1443 (17B active /397B total) Same reasoning strength at a fraction of the active parameters. Arena.ai @arena Inkling-Small ( @thinkymachines ) debuts at rank ~ #88 (1431 pts, AutoEval) in Text Arena and ~ #21 among open. This makes it one of only five US models in the top 25 open models overall. Note: this is an early AutoEval score, in which a Reward Model trained on Arena's human preference data casts automatic votes in place of live voting, giving a day-1 signal right at release. Scores are tagged preliminary until correlated with live votes. See the thread for more on how AutoEval works and the methodology behind it. Congrats to @thinkymachines for another strong contribution to the open ecosystem! We’ll continue to track to see where it lands in rank as live votes come in. 🔗 View Quoted Tweet 💬 1 🔄 3 ❤️ 8 👀 1967 📊 2 ⚡