Releases·Model Release
Moonshot AI's Kimi K3, first open model at 2.8T parameters, launches on Amazon Bedrock
AWS announced that Moonshot AI's Kimi K3 is now available on Amazon Bedrock for coding and knowledge work. Moonshot says K3 is its most capable model, the first open model to reach 2.8 trillion parameters, with native vision, a 1M-token context window, and ~2.5x better scaling efficiency than Kimi K2. It is the first open-weight model on Bedrock to support explicit prompt caching.
1010 engagement2 sourcesFri, Sep 18, 2026, 04:52 PM
AI Overview
- Kimi K3 reaches 2.8 trillion parameters, offers native vision and a 1M-token context window, with ~2.5x scaling-efficiency gain over Kimi K2 per Moonshot AI
- First open-weight model on Bedrock to support explicit prompt caching, cutting latency and input costs for reused context
- Available via us. and global. cross-Region inference profiles; global profile costs ~10% less and routes to any commercial AWS Region
- Bedrock has added dozens of open-weight models since 2025 (DeepSeek, Google, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI, Qwen) plus tool calling, streaming, and Responses/Chat Completions APIs in 2026
- Works in coding agents like OpenCode and Hermes Agent through Bedrock's Converse API and OpenAI-compatible endpoints
- Zero data retention and zero operator access apply; data stays within the AWS boundary and is not shared with Moonshot AI or used for training
Related
Releases·Model Release
Moonshot AI releases Kimi K3 model with permissive license
Moonshot AI has released the Kimi K3 language model under a license that defines 'Model as a Service' broadly, restricting third-party inference or fine-tuning services unless aggregate revenue exceeds a threshold. The license aims to balance openness with commercial protection.
Moonshot AI has released the Kimi K3 language model under a license that defines 'Model as a Service' broadly, restricting third-party inference or fine-tuning services unless aggregate revenue exceeds a threshold. The license aims to balance openness with commercial protection.
37 engagement1 sourceredditMon, Jul 27, 2026, 04:29 PM
Releases·Model Release
Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open-weight model
Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, on Hugging Face at midnight UTC on July 27, 2026. The model reportedly narrows the frontier gap to 4 points, marking a significant advance in open-weight AI.
Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, on Hugging Face at midnight UTC on July 27, 2026. The model reportedly narrows the frontier gap to 4 points, marking a significant advance in open-weight AI.
0 engagement1 sourcerssMon, Jul 27, 2026, 12:00 AM
Social·Community
Community explains Kimi K3's 2.8T-parameter MoE efficiency with MLA and KDA
A Reddit user wrote a plain-English explainer of how Kimi K3 achieves 2.8 trillion total parameters while remaining affordable to run, covering Mixture-of-Experts (MoE), KV cache, Multi-head Latent Attention (MLA), and Key-Value Decomposed Attention (KDA). The post highlights the trend toward sparser MoE models, noting Kimi K3's claimed 16-of-896 experts (about 1.8% of experts active), and compares it to earlier models like Mixtral, DeepSeek-V3, Llama 4 Maverick, and gpt-oss-120b.
A Reddit user wrote a plain-English explainer of how Kimi K3 achieves 2.8 trillion total parameters while remaining affordable to run, covering Mixture-of-Experts (MoE), KV cache, Multi-head Latent Attention (MLA), and Key-Value Decomposed Attention (KDA). The post highlights the trend toward sparser MoE models, noting Kimi K3's claimed 16-of-896 experts (about 1.8% of experts active), and compares it to earlier models like Mixtral, DeepSeek-V3, Llama 4 Maverick, and gpt-oss-120b.
4 engagement1 sourceredditSun, Jul 26, 2026, 11:11 PM
News·Markets
Exclusive: Paying for frontier AI models buys 4-month head start at 5x the cost
Mozilla's State of Open Source AI report (September 15) finds the gap between US closed frontier models and the best Chinese open-weights models has narrowed to 4.4 months. Moonshot AI's Kimi K3 scores within three points of Anthropic's Fable 5 on the Artificial Analysis Intelligence Index while costing roughly 30% as much. Mozilla's CTO argues closed models' premium applies only to specific workloads, and most organizations should default to open models.
Mozilla's State of Open Source AI report (September 15) finds the gap between US closed frontier models and the best Chinese open-weights models has narrowed to 4.4 months. Moonshot AI's Kimi K3 scores within three points of Anthropic's Fable 5 on the Artificial Analysis Intelligence Index while costing roughly 30% as much. Mozilla's CTO argues closed models' premium applies only to specific workloads, and most organizations should default to open models.
0 engagement1 sourcerssTue, Sep 15, 2026, 12:00 PM
Social·Community
Speed-up Kimi K3(2.8T) on a 16x GB10 Cluster — 30 t/s coding throughput, 136 t/s concurrency peak.
A user reports running Moonshot AI's full Kimi K3 (2.8T parameters) across a 16-node GB10 cluster, sustaining ~30 tok/s decode (peaking ~38) on coding tasks with ~750–910 tok/s prefill. The setup required custom runtime patches, modified NCCL topology, and dual MikroTik switches. Patches, configs, and build scripts were published on GitHub, offering a reference for practitioners attempting large local deployments.
A user reports running Moonshot AI's full Kimi K3 (2.8T parameters) across a 16-node GB10 cluster, sustaining ~30 tok/s decode (peaking ~38) on coding tasks with ~750–910 tok/s prefill. The setup required custom runtime patches, modified NCCL topology, and dual MikroTik switches. Patches, configs, and build scripts were published on GitHub, offering a reference for practitioners attempting large local deployments.
109 engagement1 sourceredditSun, Sep 20, 2026, 09:14 PM