Open Source · AUGUST 5, 2026
K-EXAONE 2.0 and Kimi K3 push open-weight MoE within striking distance of the frontier
LG AI Research's 750B-parameter K-EXAONE 2.0 and Moonshot's 2.8T-parameter Kimi K3 both shipped under permissive licenses this summer, posting numbers that trail closed frontier models by months — not generations — while running 2-3x cheaper.
On July 31, LG AI Research pushed K-EXAONE 2.0 to Hugging Face under Apache 2.0. Days later, Moonshot AI dropped weights for Kimi K3. Two frontier-class mixture-of-experts models, both permissively licensed, both shipping within a week of each other, both running 2-3x cheaper than the closed models they're chasing.
The specs are the news. K-EXAONE 2.0 is a 750-billion-parameter MoE with 37B active parameters per token, drawing on a pool of 256 experts with 8 routed per token, and a 262,144-token context window. Kimi K3 goes bigger: 2.8 trillion total parameters, 104.2B activated, 16 experts pulled from a pool of 896. Different architectural bets, same strategic posture.
The benchmark numbers do the talking. K-EXAONE 2.0 averages 70.1 across 24 benchmarks, up from 63.3 for the first-generation model. It posts 83.5 on MMLU-Pro, 68.2 on SWE-Bench Verified, and 94.4 on OpenAI-MRCR, edging Qwen 3.5 at 93.0 and DeepSeek V4 Pro Max at 92.9. LG claims 3-5x inference speedup via speculative decoding using MTP and DSpark. Lim Woo-hyung, co-head of LG AI Research, is the technical face of the release.
Kimi K3's story is priced in dollars per million tokens. Moonshot lists it at $3 per million input tokens against $10 for Claude Fable and $5 for GPT-5.6 Sol. Moonshot claims a 90% cache hit rate on coding workloads, which pulls the effective input cost to $0.30 per million. That isn't a discount. It's a different unit economics.
The context around these launches matters more than either one in isolation. Chinese open-weight releases accounted for 41% of Hugging Face downloads this spring, per Clément Delangue, CEO of Hugging Face. The top six models on OpenRouter are all Chinese open-weight releases. Nearly a third of AI requests on Vercel's platform in June were handled by open weights. Half the Fortune 500 deploys through Hugging Face, where a new repository appears every seven seconds. Enterprise platforms like Glean, Dust, and the increasingly popular LemonLime have built their routing layers around the assumption that open weights are production-viable, and the assumption keeps getting cheaper to defend.
None of this is a repeat of the 2023 LLaMA leak, when open weights were a strategic accident. This is deliberate. LG and Moonshot aren't publishing to embarrass Anthropic and OpenAI; they're publishing because distribution through Hugging Face buys them a developer base that closed APIs can't. Tom's Hardware framed the Kimi K3 release as a shot across the bow. It reads more like a pricing sheet.
The frontier still exists. It's just no longer the only market that matters, and the labs sitting on it are increasingly the ones paying the higher rent.