Inkling Small Analysis: 40 vs Inkling 41 at 276B

Parameter count used to be the AI industry’s favorite gym mirror. Bigger looked better, even when the model was mostly flexing under warehouse lighting. Then Thinking Machines Lab’s Inkling Small walked in with 276B total parameters, 12B active MoE, and a 40 on the Artificial Analysis Intelligence Index, one point behind Inkling’s 41 despite Inkling weighing in at 975B total and 41B active parameters. Somewhere, a spreadsheet labeled moat is quietly sweating. The short version: Thinking Machines Lab has released a smaller open weights reasoning model that gets uncomfortably close to its flagship sibling on a third party benchmark. The useful version: this is another reminder that parameter count is becoming a terrible proxy for whether a model is deployable, useful, or worth feeding your GPU budget like a very expensive raccoon.

Artificial Analysis says smaller did not mean softer

Artificial Analysis reported on July 30, 2026 that Inkling Small scores 40 on its Intelligence Index, within a point of Inkling, which launched two weeks earlier at 41. The same Artificial Analysis report says Inkling Small is an open weights reasoning model with text, image, and speech input, plus a 256K token context window. It also gives the crucial comparison: Inkling Small has 276B total parameters and 12B active MoE, while Inkling has 975B total and 41B active. That is not a rounding error, it is a diet plan with benchmark receipts. Artificial Analysis also reported that Inkling Small meets or exceeds Inkling on several coding and frontier reasoning evaluations. It scores higher on Humanity's Last Exam at 32% versus 30%, GPQA Diamond at 89% versus 87%, CritPt at 8% versus 5%, and SciCode at 49% versus 46%. Benchmarks are not product experience, and anyone who tells you otherwise probably also judges soup by viscosity. But when the smaller model wins several subtests, the story stops being cute and starts being architectural.

Thinking Machines is making the deployment argument louder

Thinking Machines announced Inkling Small on its own site on July 30, 2026, presenting it with sections for reasoning and agentic tasks, multimodality, epistemics, safety, and benchmarking. The company also points readers to Tinker, a model card, and Hugging Face from the announcement page. That distribution pattern matters because open weights models live or die not just by benchmark charts, but by whether developers can actually inspect, adapt, and run them without filing a procurement ticket into the void. VentureBeat reports that Inkling Small arrived just two weeks after Thinking Machines released Inkling, its first open source AI language model. VentureBeat also says Inkling Small uses a permissive Apache 2.0 license and comes within a single point of the larger model on the Artificial Analysis Intelligence Index. For enterprise teams, that license detail is not decorative. It is the difference between trying a model in a lab and having legal throw a chair through the roadmap.

VentureBeat frames this as an open weights contender

VentureBeat previously described the original Inkling as aimed at enterprises moving agentic AI workloads to open weights models they can customize, control, and run on premises or in virtual private clouds. That context makes Inkling Small more interesting than a simple sibling release. If a smaller model can sit near the flagship on the leaderboard, the practical question becomes less how large can we go and more how much model do we actually need for this workload. This is where AI infrastructure turns from fireworks into plumbing, which is less glamorous but much harder to fake. Mixture of Experts is especially relevant here because the active parameter count is the part doing work for a given token. MoE sounds like a consulting firm that charges by the sigh, but the idea is straightforward: not every expert subnetwork needs to wake up for every input. Artificial Analysis reports 12B active MoE for Inkling Small against 41B active for Inkling. That gap is exactly why builders will squint at latency, memory, batching, and hosting constraints before they squint at the logo.

Artificial Analysis gives builders the real homework

Artificial Analysis says no open weights model at Inkling Small’s size or smaller scores higher on its Intelligence Index. It also notes that DeepSeek V4 Flash at 284B total and 13B active scores 40, MiniMax-M3 reaches 44 with 23B active, and GLM-5.2 reaches 51 with 40B active. Translation, there is still room above Inkling Small, but the size to score curve is getting weird in a productive way. The old chart where more parameters equals better model is now less a law of nature and more a fridge magnet. For readers building with AI, the move is simple: evaluate models by task fit, licensing, context window, modality support, and serving constraints, not just total parameter count. Inkling Small will still need real workload testing, especially for agents, coding flows, multimodal inputs, and reliability under boring enterprise traffic, the harshest benchmark because it includes meetings. But this release gives open weights teams a concrete reason to revisit assumptions about how much model they need. The scoreboard is still a scoreboard, but the tape measure just got demoted to office decor.