Kimi K2 Thinking’s benchmark performance

Hey there!

Welcome back to The Pulse, where we dive into interesting AI stories and trends backed by data, all presented through simple visuals.

> Kimi K2 Thinking: first open model to match GPT-5 & Claude 4.5 on core reasoning + coding tasks

> ranks #2 on Artificial Analysis Intelligence Index, just behind GPT-5 (high)

> beats GPT-5 & Claude 4.5 on Humanity’s Last Exam (51 vs 42 top) + BrowseComp (60.2 vs 54.9 top) in strongest agentic showing yet

> 71.3 on SWE-Bench (coding) + 84.6 on MMLU Pro = deep reasoning + broad knowledge coverage

> trained for just $4.6M vs hundreds of millions for closed models

> inference runs ~6× cheaper than Claude & ~10× cheaper than GPT-5 while matching the reasoning range

> MoE (~32B active params) architecture makes it light, fast, & scalable for enterprise deployment too

> planned over next 7 years:

• 11+ GW in datacenter expansion

• $310B+ in cloud usage (additive to capex)

• $1T+ in circulation ($1.4T in commitment overall, thru 8 yrs acc. to Altman)

> doubling down on compute expansion for AI scaling; to build multi-GW spine across regions + layer multi-cloud & specialized cloud lanes

> "infinite money glitch" created through investments feeding back as revenue, creating loop that grows valuations & works only if customer reach → paid usage

> vendor financing is normal; fine if demand covers cash flows

> slowly putting into existence Altman's eventual hope of 1 GW (~750K homes worth of energy) added per week

> Q3 2025 revenue hit $1.36B, up ~133% YoY, beating estimates (~$1.28B)

> net loss narrowed to $0.22/share vs. $1.82/share in Q3 2024; adjusted loss just $0.08/share (vs -$0.36 est.)

> lowered 2025 revenue guidance to $5.05-$5.15B (from $5.15–$5.35B) due to data center delay, so shares dropped ~6%

> 2025 capex forecast at $12-14B, expected to double in 2026

> signed $14B deal with Meta + $6.5B with OpenAI, pushing revenue backlog past $55B

> CoreWeave, Lambda, Nscale, Nebius in defining a new cloud category: sovereign AI infra built around GPU supply

> signal: record top-line growth + deferred profits = early hyperscaler playbook for neo-clouds