AWS Adds Open Weight Models to Amazon Bedrock for Terminal-Based Coding Agents via OpenCode

Amazon Web Services published a technical guide showing how to run the open source coding agent OpenCode against open weight large language models hosted on Amazon Bedrock, positioning the combination as an alternative to proprietary API-based coding assistants that require sending code to third-party services.

OpenCode is a terminal-native coding agent built in Go that reads and edits files, executes shell commands, and parses project structure through Language Server Protocol diagnostics. It connects to more than 75 LLM providers. AWS's guide focuses specifically on pairing it with three open weight models available on Bedrock: Moonshot AI's Kimi K3, OpenAI's GPT-OSS 120B, and NVIDIA's Nemotron 3 Super 120B.

Key technical details

Kimi K3 supports a 1 million-token context window — enough, according to AWS, to load an entire code repository rather than a handful of files for cross-file reasoning. The model includes an adjustable reasoning depth parameter (reasoning_config, set to low, high, or max) that trades latency for accuracy on harder problems.

Nemotron 3 Super 120B uses a Mixture-of-Experts architecture that activates only 12 billion of its 120 billion total parameters per token. NVIDIA claims this delivers up to 7x higher throughput compared to dense models of similar size, making it suited to high-volume, latency-sensitive tasks like code completion.

GPT-OSS 120B is positioned as a general-purpose generation model for full-service code writing and multi-file implementations.

Bedrock offers three pricing tiers for these models: Priority for latency-sensitive production use, Standard for on-demand pay-per-token inference, and Flex, which AWS says costs 50% less for workloads that can tolerate variable latency. A global cross-Region inference profile for Kimi K3 (global.moonshotai.kimi-k3) costs approximately 10% less than a geography-locked profile, while a US-specific profile (us.moonshotai.kimi-k3) keeps processing within US borders for data residency needs. Default account limits are 100 million tokens per minute and 10,000 requests per minute.

Benchmark claims cited

AWS cites a CrowdStrike case study in which a fine-tuned NVIDIA Nemotron model reportedly achieved 96% valid query accuracy on an internal task, compared to 61% for GPT-4o and 94% for Claude Sonnet 4.5. These figures come from CrowdStrike and NVIDIA, not independent benchmarking, and reflect a narrow, domain-specific evaluation rather than general coding ability.

AWS also references a McKinsey 2025 report claiming 76% of organizations expect to increase open source AI usage, and a Gartner 2026 analysis stating agentic workflows can multiply token consumption 5–30x compared to single-turn queries — a dynamic that makes per-token pricing differences between models more consequential at scale.

Architecture

OpenCode runs locally in a terminal and calls the Bedrock Converse API for inference; Bedrock hosts the models as managed, serverless endpoints. Requests are authenticated via AWS IAM and logged through AWS CloudTrail. Bedrock inherits existing AWS security controls including PrivateLink connectivity and encryption, and AWS states it does not use customer inputs or outputs to train foundation models. Bedrock is in scope for HIPAA, SOC 2, ISO 27001, FedRAMP, and GDPR compliance programs, according to AWS.

OpenCode's agent design allows different models to be assigned to different roles within one session — for example, a reasoning model like Kimi K3 for planning and a faster model like Nemotron for generation — creating a multi-model workflow without switching tools.

AWS says Ethara.AI uses this architecture in production with multi-agent orchestration for engineering and research workflows, though no performance or cost figures from that deployment were disclosed.

What this means

This is a packaging and integration announcement, not a new model launch. AWS isn't shipping a new checkpoint — it's making existing open weight models (from Moonshot AI, OpenAI, and NVIDIA) easier to use inside a coding agent without leaving the AWS compliance boundary. The pitch is aimed squarely at enterprises with data residency constraints or high-volume agentic workloads where per-token costs on frontier proprietary APIs add up quickly under Gartner's cited 5–30x token multiplier for agentic loops.

The real signal here is Bedrock's growing role as a neutral hosting layer for open weight models, letting teams switch between Kimi K3, GPT-OSS 120B, and Nemotron with a single API parameter rather than re-architecting their tooling. Whether the cost savings materialize depends heavily on workload pattern — Flex tier's 50% discount only helps if latency tolerance is real, and the CrowdStrike benchmark cited is task-specific rather than a general claim about open weight models beating GPT-4o or Claude on coding.