Short version: Thinking Machines Lab released Inkling-Small on July 31 — a 276B open-weight MoE that scores 40 on the Artificial Analysis Intelligence Index against the 975B flagship's 41, and reportedly beats that flagship on the agentic and coding benchmarks that matter for building products. But the number to build your plan around is the ~12B active parameters, because that's what sets your inference bill. Here's the split, and what a team of one should actually do with it.
What shipped#
On July 31, 2026, Thinking Machines Lab — the lab founded by former OpenAI CTO Mira Murati — released Inkling-Small, the smaller sibling of the 975B-parameter Inkling flagship it opened on July 15. The specs, from the model card and launch coverage:
- 276B total parameters, mixture-of-experts, with ~12B active per token.
- Multimodal input — text, image, and audio — to text output.
- 1M-token context, Apache 2.0 license, full weights on Hugging Face.
- Fine-tuning offered through Thinking Machines' Tinker API for teams that don't want to run training themselves.
It scores 40 on the Artificial Analysis Intelligence Index — one point under the 975B flagship's 41 — at under a third of the total size. And the more interesting claim from the launch write-ups: on several agentic and coding evaluations (HLE, SWE-bench Verified, Terminal-Bench 2.1, ARC-AGI-2), Inkling-Small actually beats the larger Inkling. Treat those specific wins as reported until you've run your own eval, but the direction is clear — for the tasks founders build on, small isn't the compromise here.
The one number that sets your bill#
The mistake is to read "276B" and mentally file this next to other giant models. Don't. In a mixture-of-experts model, only a slice of the parameters fire on any given token — here about 12B out of 276B. That slice is what your GPU actually computes each step, so:
- Cost-per-token and latency track the ~12B active count. At inference time, Inkling-Small behaves like a ~12B model. That's cheap and fast.
- VRAM tracks the full 276B. You still have to hold every expert in memory, because you don't know in advance which experts a token will route to. At 4-bit that's roughly 140GB — a small multi-GPU box, not a single card.
So the correct way to plan is split-brain: price your per-request economics off 12B, and size your hardware off 276B. Founders who conflate the two either over-provision compute they'll never use or under-provision memory and can't load the model at all.
What it means: if you were priced out of a frontier API on volume, a 12B-active open weight that lands within a point of a 975B model is exactly the kind of thing that flips your unit economics — provided you can keep a two-GPU box busy enough to beat the hosted rate.
Own-and-fine-tune, not a cheaper API call#
The reflex is to treat every new cheap model as "a lower number on the price card." Inkling-Small is a different category: Apache 2.0 means ownership. You can fine-tune it on your proprietary data, serve it on infrastructure no vendor can read or revoke, and pin a version that will never be deprecated out from under you at the end of a support window. That's a strategic asset for anything the model sits at the center of — a coding agent, a support brain, a document pipeline.
The trade is operational burden for control. It's worth taking when the model is core to your product and you can keep the hardware utilized; it's not worth it for occasional calls, where a hosted endpoint's zero idle cost wins. This is the same rent-vs-own decision the open-weight wave has forced all quarter — we walked the full hardware math on self-hosting Kimi K3, and the open-weight bet between Kimi K3 and Inkling frames the same choice at the model layer.
What to do this week#
- Run your own eval, not the leaderboard. The "beats the flagship on coding" claim is promising but it's launch-week reporting. Point Inkling-Small at your hardest real task class and compare cost-per-completed-task against your current model.
- Do the memory math before you fantasize about a self-host. ~140GB at 4-bit. Confirm your box can hold all 276B before you plan an inference budget around the 12B compute profile.
- If the model is core, pilot a fine-tune. The whole point of an Apache 2.0 open weight is specialization. Use Tinker or your own stack to tune it on your domain and measure the lift against the base — that's where an owned model pulls ahead of a rented one.
- If it's peripheral, stay on a hosted endpoint. Don't take on GPU ops to save pennies on calls you make a few thousand times a day. Ownership is a tool for your core loop, not your whole stack.
