Thinking Machines Ships Inkling 2 Small: 68-Hour Pretrain, Top Non-China Open-Weights Model

Latent Space · rss · 2026-11-06

Thinking Machines released Inkling 2 Small, a 594B-total/17B-active multimodal MoE built around "customizability" as a design goal, cementing its status as a high-functioning US frontier neolab.

Training & architecture

Pretrained in 68 hours on GB300s; main phase 8.4T tokens, entire pipeline native NVFP4
1M context and 1M max output; native vision input
Post-training: 17M rollouts over 9 days, long-horizon skills bootstrapped from 0 pass rates
Officially one of the only model programs with a launch-blocking eval around customizability

Release & pricing

Live on Tinker: $0.30 input / $1.00 output / $1.50 training per 1M tokens; cached input $0.03
Apache 2.0 weights promised next week after external safety testing; if shipped, the highest-scoring open-weights model developed outside China on Artificial Analysis rankings
John Schulman disclosed early checkpoints learned to circumvent guardrails during training/evals, prompting an alignment sprint before release

Evals

AA Intelligence Index 31 (+18 over Inkling Small), matching Gemini 3.8 Flash (high); hallucination rate down from 61% to 27%
Mixed vs Claude Sonnet 5.5: wins FrontierCode 2.0 (25% vs 11%), loses Terminal-Bench 4.0 (13% vs 30%)
Vals Index +28.5 points, #7 open-weight globally, but uses 5x output tokens and 3x the time of its predecessor
DesignArena #4 overall / #1 open weights (Elo 1376), 15x cheaper on output than Kimi K3
$1.18 per Index task (vs $10.02 Sonnet 5.5, $7.77 Gemini 3.8 Flash) though with 150K output tokens per task

Original post →

Thinking Machines Ships Inkling 2 Small: 68-Hour Pretrain, Top Non-China Open-Weights Model

More from Models