Moonshot AI's 2.8 Trillion-Parameter Kimi K3 Launches on Amazon Bedrock with 1M-Token Context
Moonshot AI's Kimi K3, described by the company as the first open model to reach 2.8 trillion parameters, is now available on Amazon Bedrock. It features native vision, a 1-million-token context window, and is the first open-weight model on Bedrock to support explicit prompt caching.
What happened
Amazon Web Services has added Kimi K3, Moonshot AI's newest open-weight model, to Amazon Bedrock. According to Moonshot AI, Kimi K3 is its most capable model to date and the first open model to reach 2.8 trillion parameters. The model ships with native vision capabilities and a 1-million-token context window, and Moonshot AI claims it delivers roughly 2.5x better scaling efficiency than its predecessor, Kimi K2.
None of Moonshot AI's parameter count, efficiency, or capability claims have been independently verified through published benchmark scores in this release. AWS's blog post does not include third-party benchmark comparisons (e.g., MMLU, HumanEval, or SWE-bench figures) for Kimi K3.
Availability and technical details
Kimi K3 is accessible through the Amazon Bedrock console (Test > Playground) or programmatically via the bedrock-runtime endpoint, which supports OpenAI-compatible Responses and Chat Completions APIs alongside Bedrock's native Invoke and Converse APIs. Two cross-Region inference profiles are available at launch: a global profile (global.moonshotai.kimi-k3) that routes requests to any supported commercial AWS Region and costs approximately 10% less than regional routing, and a US-only profile (us.moonshotai.kimi-k3) for data residency requirements.
Pricing per million tokens was not disclosed in AWS's announcement.
AWS says Kimi K3 is the first open-weight model on Bedrock to support explicit prompt caching. Developers can mark reusable prompt prefixes (minimum 1,024 tokens) with a prompt_cache_breakpoint. Cached tokens are billed at a higher rate on write but held for at least 30 minutes; matching subsequent requests get discounted input-token pricing and are exempt from input-tokens-per-minute quota limits. Exact cache pricing was not specified.
Standard Bedrock data protections apply: processing stays within the AWS data boundary, prompts are not shared with Moonshot AI or used for training, zero data retention is enabled by default, and AWS states its own operators cannot access prompts or completions during inference.
Ecosystem integration
Kimi K3 is accessible through third-party tools that support Bedrock or OpenAI-compatible providers, including the open-source coding agent OpenCode (via a native amazon-bedrock provider using the Converse API) and Hermes Agent, an open-source productivity assistant. AWS demonstrated OpenCode building a browser-based game using the model as a proof of long-horizon coding capability.
The launch continues AWS's expansion of open-weight model support on Bedrock, which since 2025 has added models from DeepSeek, Google, MiniMax, Mistral AI, Moonshot AI, NVIDIA, OpenAI, and Qwen. In 2026, Bedrock added platform-level support for tool calling, structured output, reasoning, response streaming, and the Responses/Chat Completions APIs — capabilities that apply automatically to newly onboarded open-weight models like Kimi K3.
What this means
Kimi K3's headline figure — 2.8 trillion parameters — would make it one of the largest open-weight models publicly available, but AWS's post relies entirely on Moonshot AI's own characterization without independent benchmark data to substantiate capability or efficiency claims. The more concrete news is infrastructure: Bedrock's explicit prompt caching support for an open-weight model signals AWS is treating Moonshot AI's models as first-class citizens on its platform, not just a compatibility checkbox. For enterprises evaluating large open-weight models for long-context coding and document workflows, the combination of a 1M-token window, native vision, and cache-based cost controls is the practical draw — pending real-world benchmarking and pricing transparency that AWS has not yet published.
Related Articles
DeepSeek Ships V4.1-Flash With Novel Encoder-Decoder Architecture, Cuts KV Cache to 1/8 of Predecessor
DeepSeek released V4.1-Flash, a 763B-parameter model built on a new causal encoder-decoder architecture that splits 8B active parameters for prefill and 16B for decode. The model adds native vision support, a 1M-token context window, and shrinks KV cache footprint to roughly 1/8 of DeepSeek V4 Flash, while retiring V4 Pro.
Z.ai Releases GLM-5.3-FlashX, a 200 Tokens/Second Variant of Its GLM-5.3-Flash Model
Z.ai has released GLM-5.3-FlashX, a high-speed variant of GLM-5.3-Flash built on a hybrid sparse and linear attention architecture with 320B total parameters (18B active). The model supports a 1M-token context window and claims inference speeds of up to 200 tokens per second.
Shanghai AI Lab Releases Atria Dawn Preview, a 744B-Parameter MoE Agentic Model Built on GLM-5.2
Shanghai Artificial Intelligence Laboratory has released Atria Dawn Preview, a text-only agentic model built on the 744B-parameter MoE GLM-5.2 foundation model with a 256K context window. The model targets multi-step research, coding, and productivity tasks, with benchmark results claimed to compete with DeepSeek V4 Pro, Kimi K3, and Claude Opus 5.
China Telecom Releases Xing4.0-29B-A4B, a 29B MoE Model Trained Entirely on Ascend NPUs
China Telecom Artificial Intelligence Technology has released Xing4.0-29B-A4B, a 29-billion-parameter mixture-of-experts model with only 4B parameters active per token and native 256K context. The company claims it is the first model of this scale trained entirely on Huawei's Ascend NPU platform using the MindSpore framework.
Comments
Loading...