What AWS described
AWS's post presents OpenCode as a terminal-native, open-source coding agent written in Go that reads and edits files, runs shell commands and uses LSP diagnostics, connecting to 75-plus LLM providers. It proposes running it with open weight models on Bedrock so inference stays inside the AWS account, switching models as a single API parameter change. The post names Kimi K3, OpenAI GPT-OSS 120B and NVIDIA Nemotron 3 Super 120B as examples and says it will walk through setup and multi-model workflows. [1]
Routing and cost claims
For regional control, the post says Kimi K3 can be invoked via a cross-Region inference profile, with a global profile routing requests to any supported commercial AWS Region and costing roughly 10 percent less than a geographic profile; a US profile keeps processing within the US geography. AWS says models accessed through an in-Region or geographic profile run in that Region or geography. [1]
Vendor-cited performance, as claimed
The post cites a CrowdStrike fine-tune of NVIDIA Nemotron reaching 96 percent valid query accuracy, versus 61 percent for GPT-4o and 94 percent for Claude Sonnet 4.5, as an example of domain-specific open weight performance. These are figures AWS attributes to CrowdStrike, not independent AWS benchmarking. AWS also cites a McKinsey 2025 report that 76 percent of organizations expect to increase open source AI usage. [1]
Deployment and incentives
AWS says Ethara.AI already runs this architecture in production with multi-agent orchestration. It argues token consumption in agentic workflows makes cost-per-token critical, and that open weights allow fine-tuning and domain adaptation so smaller models can replace pricier general-purpose ones. [1]
Sources
- Use open weight models as your AI coding agent with Amazon Bedrock
AWS Machine Learning Blog · Primary source ·
“Use open weight models as your AI coding agent with Amazon Bedrock”
“OpenCode is an open source, terminal-native AI coding agent built in Go.”