Artificial Intelligence: Citi‘s Inference Ahead - Open Reaches the Frontier

Research evidence excerpt

Artificial Intelligence: Citi‘s Inference Ahead - Open Reaches the Frontier

Viewpoint |

19 Jul 2026 11:12:31 ET │ 18 pages

Artificial Intelligence

Citi’s Inference Ahead - Open Reaches the Frontier

CITI'S TAKE

Heath Terry AC

Open-weight models continue to narrow the gap (Figure 1), with +1-212-723-4624

Moonshot's Kimi K3 score of 57 ranking close to the top proprietary model’s heath.terry@citi.com

score of 60 on AA’s Intelligence Index (Figure 2), furthering concerns over the

sustainable economics of the AI Labs, even as the cost of intelligence Shelby Spencer

continues to decline (Figure 7). Importantly, model performance gains are +1-212-816-0416

still outpacing rising output token costs, particularly at the frontier. Higher shelby.spencer@citi.com

costs will continue to encourage efforts at routing to more efficient (often

open) models, driving enterprise adoption and infrastructure needs – Ashley Kim

something we expect to be reflected as hyperscalers report accelerating AI +1-212-816-6689

revenues, growing cloud backlogs, and rising CapEx. ashley.kim@citi.com

Janna WithrowNew Models. Kimi K3, released on Wednesday, is the largest open-weight model to

date at 2.8T parameters, with weights to be released by July 27 (Moonshot AI, 7/16). +1-212-723-0439

This model raises the floor for open intelligence just as Gemini 3.5 Pro reportedly janna.withrow@citi.com

sees delays after falling short of internal goals (Bloomberg, 7/16). Thinking Machines

Lab's first open-weights model, Inkling, is among the first GB300-trained models

(7/15) and sits 10 points below GLM-5.2 at more than 6x DeepSeek's blended price

(Figure 2). Its MoE design follows DeepSeek-V3, and it leveraged synthetic data from

open-weight models like Kimi K2.5 for post-training. Those interdependencies