Kimi K3 is a large open model from Moonshot AI with roughly 2.8T–3T parameters and about 1.45 TB of expert weights. Recent coverage has focused on deployment: Cerebras spreads the model across multiple WSE-3T wafers with 4-bit quantization to reach 1,000–2,000+ tokens per second, AWS added Kimi K3 to Amazon Bedrock, and Cognition post-trained its SWE-2 coding model from it. Moonshot AI aims for $2 billion in annualized revenue in 2026 using Kimi K3.
Cerebras deploys 3T-parameter Kimi K3 on wafer-scale chipsCerebras explained how it splits the open-source ~3T-parameter Kimi K3 across multiple WSE-3T wafers (44GB on-chip memory each) using 4-bit quantization to reach 1,000–2,000+ tokens per second. Kimi K3 scores 88.3% on Terminal-Bench 2.1 vs GPT-5.6 Sol's 88.8%, and 91.2% on BrowseComp vs 90.4%.
Survey finds leading AI models converge on the same "controversial" take about humansAsked repeatedly for their most controversial take on humans, all 12 models from ten labs in three countries gave at least once the same view: people adopt their group's beliefs and then invent reasons; 7 of 12 said it every time. Claude Opus 5.5 once argued the reverse, and Grok 4.7 once declined to answer.
ThinkingBox benchmark: one-shot agent success rarely repeatsMicrosoft authors released ThinkingBox-Bench: 507 business workflows across 5 domains, each run 20 times and graded on terminal backend state. Kimi-K3 solved 93.89% of tasks at least once but only 13.41% all 20 times, while Claude Opus 5 discovered fewer (79.09%) yet repeated far more (47.53%), showing discovery and repeatability rank models almost oppositely.
China, U.S. Must Manage AI Risks TogetherFudan scholar Zhao Minghao argues AI challenges are already real, and viewing AI only through competition could cost China and the U.S. chances to coordinate on shared risks. He notes the Trump administration is rolling out a broad AI strategy and says the U.S. must win the race with China, while Chinese open-source models gain global ground on performance.
Kimi K3 Now Available on Amazon BedrockAWS announced that the Kimi K3 model is now available on Amazon Bedrock, letting enterprise users call it directly to build AI applications. Pricing and regional availability were not disclosed in the announcement text.
Report: China narrows AI gap with U.S.The report says China has been rapidly closing its AI lead over the U.S. in model capabilities and research despite U.S. export restrictions on advanced chips and equipment. Experts describe the U.S.–China model gap as narrowing, noting Chinese efforts such as Moonshot AI’s Kimi K3.
Open Chinese models narrow gap with frontier modelsA Mozilla report finds that the performance gap between Chinese open-weight models and US frontier models has narrowed to roughly 4.4 months. It highlights Moonshot AI's Kimi K3 scoring three points behind Anthropic's Fable 5 on a composite index while costing about 30% as much, suggesting open models are preferable for most routine workloads.
CrofAI exposed as OpenRouter wrapper with big markupsResearchers allege CrofAI (NahCrofAI) was an OpenRouter wrapper that silently routed requests to cheaper/weaker models while charging up to 20x markups. The operator denied then backtracked and ultimately wiped the service; examples include routing Kimi K3 requests to GLM models, per the exposé.
Do tech reports matter for PhD appsA Reddit post asks whether tech reports for large models (not arXiv papers) — e.g. Kimi K3, DeepSeek, Gemini, Mistral Leanstral — carry similar weight to a first‑author A* paper in PhD applications, seeking community input.
Terminal Bench v4 ranks GLM-5.3 topA shared Terminal Bench v4 ranking shows GLM-5.3 leading at 41.9%, followed by GLM-5.3-Flash (32.8%) and DSV4.1-Flash (26.8%). Qwen3.8-27B is the only smaller model with notable results. The list reflects community perceptions of open vs closed models.
Moonshot Aims for $2B Annualized Sales in 2026Moonshot AI aims to lift annualized revenue to $2 billion by year-end leveraging its Kimi K3 model, intensifying competition with peers like Anthropic and Z.AI.
Cognition Unveils SWE-2 Coding ModelCognition launched SWE-2, a coding model scoring 50.0 on FrontierCode 1.1 Main—within one point of Fable 5.1 while 64% cheaper; it was post-trained from the 2.8T-parameter Kimi K3 and uses single-run RL across reasoning levels to improve cost-performance.
Kimi K3 runs at ~1 token/s on MacBook Pro with 4 SSDsA report shows Kimi K3 (2.8T parameters, 1.45 TB of expert weights) running on an M5 Max MacBook Pro (128 GB) with experts streamed from four SSDs; a 512-token answer achieved ~1.00 token/s steady decode. Measurement date is 2026-09-08. Related projects mentioned are Deltafin and ARGODRIVE.
The A.I. Daily