A top White House official is escalating the fight over Moonshot AI's viral Kimi K3 model

tech

A top White House official is escalating the fight over Moonshot AI's viral Kimi K3 model

"We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," a top White House official wrote on X.

TL;DR

A top White House official accused Moonshot AI of stealing from Anthropic to develop its Kimi K3 model.
Michael Kratsios stated Moonshot AI used a sophisticated internal platform to conduct large-scale distillation against U.S. models to avoid detection.
Kimi K3 is a highly regarded open-source AI model that performs comparably to leading U.S. models, despite expected free availability.
Legitimate AI distillation for creating smaller, efficient models is acceptable, but covert industrial distillation is deemed unacceptable.
This statement aligns with Treasury Secretary Scott Bessent's suggestion of potential sanctions on Chinese companies for improperly trained models.
Anthropic has previously accused Moonshot AI and others of using distillation attacks that violate their terms of service.
Anthropic's head of public policy views illicit distillation as IP theft and industrial espionage that poses national security risks.

Continua a leggere l'articolo originale

Kimi.ai reposted

Design Arena

BREAKING: Kimi K3 by @Kimi_Moonshot is 1st overall on 3D Design with an Elo of 1450.

This is a 6 position and 108 Elo jump from @Kimi_Moonshot's previous model, Kimi K2.6. This performance puts Kimi K2.6 82 Elo ahead of Claude Fable 5 by @AnthropicAI in 2nd and 87 Elo ahead of GLM 5.2 by @Zai_org in 3rd.

Congratulations to the @Kimi_Moonshot team on this accomplishment!

Kimi.ai reposted

Artificial Analysis

Kimi K3 is second only to Fable 5 on AA-Briefcase, our agentic knowledge work benchmark, but costs more than Opus 4.8 to run while averaging nearly an hour per task

Last week @Kimi_Moonshot released Kimi K3, a 2.8T parameter model that scores 57 on the Artificial Analysis Intelligence Index, comparable to models such as Opus 4.8 and GPT-5.5. On AA-Briefcase, Kimi K3 scores an Elo of 1543, a +727 improvement over Kimi K2.6 and the second highest score recorded, behind only Claude Fable 5 (1574)

AA-Briefcase is our new proprietary benchmark for agentic knowledge work, testing models on a fully private dataset of realistic tasks across thousands of complex input files. Tasks require deliverables such as spreadsheets, presentations, and UI mock-ups, with performance combined into a single AA-Briefcase Elo based on correctness, analytical quality, and presentation quality

Key results for Kimi K3 on AA-Briefcase:

➤ Second only to Fable 5: Kimi K3 achieves an AA-Briefcase Elo of 1543, the second-highest score overall, ahead of GPT-5.6 Sol (max, 1501), Claude Sonnet 5 (max, 1388), and Claude Opus 4.8 (max, 1347). This is a +727 improvement over the previous-generation Kimi K2.6 (816) and puts Kimi K3 only behind Fable 5

➤ Strong objective and analytical performance, with comparatively weaker presentation: Kimi K3 achieves a rubric pass rate of 51%, second only to Claude Fable 5 (56%) and ahead of Claude Sonnet 5 (max, 42.3%) and GPT-5.6 Sol (max, 41.8%). It also records an analytical quality Elo of 1754, comparable to Claude Fable 5 (1744). Presentation quality is comparatively weaker, with a Presentation Elo of 1471, below GPT-5.6 Sol (max, 1660) and Claude Opus 4.8 (max, 1492)

➤ ~10x increase in Cost per Task: Kimi K3 averages a cost of $10.57 per task, a ~10x increase from Kimi K2.6, placing it among the most expensive models to run on AA-Briefcase. This is driven by model token pricing, increased output tokens and relatively high turn use, averaging 83 turns per task, versus 67 for Claude Fable 5 and 50 for GPT-5.6 Sol (max). Kimi K3 is priced at $3/$15 per 1M input/output tokens, with a 90% discount for cached tokens

➤ Averages nearly an hour per task: Kimi K3 has an average Time per AA-Briefcase Task of 56.4 minutes. This is driven by a high number of turns, high output token use, and lower speeds using the first-party Kimi API

Yann LeCun reposted

Jitendra MALIK

Highly performant open weights frontier models such as Kimi are a competitive threat to OpenAI & Anthropic, but probably for everyone else these are a win. Hope more US entities will release top quality open weights models as well. Government regulation of AI models to prevent public harm could be done equally for proprietary or open weight models. Classic arguments in favor of open source software apply to AI models too.

Elon Musk reposted

Howard Lutnick

CAISI’s latest report shows that Kimi K3 remains behind America’s leading frontier AI models.

The United States continues to lead in frontier AI because we’re home to the greatest innovators and technologists the world has ever seen.

Kimi.ai

Happy to have @FireworksAI_HQ as our day0 launch partner and bring Kimi K3 to more developers.

With Fireworks, you can now deploy and fine-tune the 2.8T Kimi K3 model with just a few clicks! https://t.co/mwQKjvXGdl

clem 🤗 reposted

Arena.ai

In Frontend Code Arena, Kimi K3 (Max) by @Kimi_Moonshot is ranked #1 among open and #1 overall!

It’s the #1 open-weight model in all Domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Gaming, Simulations, and Content Creation Tools.

And for overall, #1 in 5 of 7 Domains, landing #2 only in Gaming and Content Creation Tools.

clem 🤗 reposted

Lewis Tunstall

We ran a large-scale distillation attack on the Kimi K3 technical report by reading it in parallel at the Hugging Face Journal Club :)

https://t.co/C5H7Ox8Jty

Our main takeaway is that there doesn't appear to be any "secret sauce" behind frontier performance.

Instead, K3 is a demonstration of how many difficult algorithmic and infrastructure decisions have to work together (distillation, RL scheduling, environments, quantisation, MoE sharding etc) to push a model to the frontier.

Here’s what we learned:

• Frontier post-training increasingly looks like expert training followed by distillation. K3 trains specialists across three domains and three reasoning-effort levels, then distils the resulting nine experts back into a single checkpoint using multi-teacher OPD

• Reasoning effort is treated as a trainable capability. Token budgets are estimated from the SFT model, and a stage-wise curriculum anneals from long to short rollouts to produce low-, high- and max-effort experts.

• Partial rollouts keep expensive RL infrastructure busy. Completed trajectories can trigger immediate updates, while unfinished ones are carried into later iterations and reprioritised through a priority queue. Per-token corrections help control the resulting off-policy behaviour.

• The reward model is itself an agent. Rather than relying only on fixed or binary rewards, it generates task-specific rubrics and scores responses against them on the fly.

• The environment interface is designed for composition. A unified, white-box abstraction allows the same training system to operate across different agent harnesses like Codex, CC, Hermes instead of specialising around a single one.

• The systems work is as interesting as the algorithmic work. GPUs dynamically switch between training and inference, rollout concurrency responds to KV-cache pressure, and reference models can live on CPU until needed.

• Quantisation is part of training—not merely deployment. K3 uses QAT for weights and activations, while matching the trainer and inference server’s quantisation schemes to reduce train–inference mismatch.

• The chat template is novel and introduces an extensible token markup language is intended to accommodate new tools and modalities without repeatedly redesigning the template.

• Synthetic task generation is becoming increasingly autonomous. Agents traverse a knowledge graph and search the web to generate training tasks at scale, although the report leaves some details of their integration underspecified. (This part was very cool)

Kimi.ai

Kimi K3 is now live on @databricks! https://t.co/wo5FGV8kzz