The refreshing thing about Thinking Machines Lab’s Inkling launch isn’t the 975B-parameter MoE or the 1M-token context. It’s that the model card leads with a line most vendors would bury: Inkling is not the strongest overall model available today, open or closed. Then it prints the table proving it — 77.6% on SWE-Bench Verified against Fable 5’s 95.0%, 63.8% on Terminal Bench against GPT 5.6’s 89.5%.

That framing tells you what this release is actually for. With 41B active out of 975B total, and an Inkling-Small at 12B active, the pitch is customization economics, not leaderboard wins. Open weights on Hugging Face plus a fine-tuning platform is a bet that most production teams don’t need the frontier — they need a competent multimodal base they can adapt and self-host without paying peak-model latency on every request.

The detail I’ll be borrowing for eval harnesses is controllable thinking effort as a first-class knob (0.2 to 0.99), with the claim that Inkling matches a stronger model on Terminal Bench at roughly a third of the tokens. In production, “which model” is often the wrong question; “how much compute per request, gated on task difficulty” is the real latency-budget lever, and baking that into the model instead of into prompt hacks is the right layer for it.

The HN thread is mostly arguing about whether “open weights” without open data counts. Fair. But the sharper operational question is whether a 41B-active model you can fine-tune on your own evals beats a frontier API you can’t touch — for your workload, not for a benchmark leaderboard. Which side of that trade is your team actually on?