Kimi K3 Launch Claims Scrutinized: Open Weights and Cost Advantages Overstated

·1 views

Moonshot AI launched its Kimi K3 model on July 16, 2026, a 2.8-trillion-parameter Mixture-of-Experts system with a one-million-token context window and native vision and tool-use capabilities. Despite widespread claims, the model's weights are not yet publicly downloadable — Moonshot has only committed to releasing them by July 27, 2026. The commonly cited figure of 2.8 trillion active parameters per token is also misleading, as the architecture routes each token through just 16 of 896 experts, with the true active parameter count undisclosed. On pricing, the API charges $3 per million input tokens and $15 per million output tokens, but independent testing found K3 generates roughly twice as many output tokens as comparable models, potentially negating its cost advantage. Analysts recommend task-level evaluation and output caps before treating K3 as a default production route.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

How to Rotate TLS Certificates in Apigee X Without Causing API Downtime

Certificate and key rotation in production API environments is a critical maintenance task that, if done incorrectly, can cause SSL handshake failures and widespread API outages. In Apigee X, the recommended approach involves uploading a new certificate alongside the existing one rather than replacing it immediately, ensuring both are valid during a transition period. Traffic is then gradually shifted to the new certificate by updating Target Servers, Environment Groups, or load balancer listeners, followed by thorough testing via curl and monitoring dashboards. Only after confirming stable operation and allowing time for DNS propagation and client cache expiration should the old certificate be removed. This overlapping strategy mirrors enterprise best practices for zero-downtime rotation and applies to TLS, mutual TLS, and Keystore or Truststore configurations within Apigee X.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

A11 Framework Claims to Fix AI Agent Failures Through Vertical Role Architecture

A new AI framework called A11 proposes a layered, vertical architecture to address common failure modes in modern AI agent systems. The framework organizes agent behavior across defined levels—S1 through S11—each with a strict, non-interchangeable purpose covering intention, values, knowledge, integration, action, and verification. A11 aims to prevent critical rules from being lost during context compression, stop role-mixing among agents, and block prompt injection by treating external data as input to be validated against fixed constraints rather than as executable instructions. When conflicts arise between task demands and stored values or rules, the system logs a TensionPoint and redirects the agent to seek clarification rather than proceed. While A11 does not make the underlying language model deterministic, it claims to impose a predictable decision cycle around the model's probabilistic outputs by anchoring key variables at each architectural level.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why JSON-First API Clients Like Postman Fall Short for Protobuf and gRPC

Most popular API clients, including Postman, were designed around REST and JSON workflows, making them poorly suited for protobuf and gRPC-based APIs. These tools struggle with schema-aware editing, as protobuf messages rely on .proto files with strict field types, imports, and structures that a plain JSON text box cannot interpret. Setting up .proto files in general-purpose clients is cumbersome because import paths and dependencies rarely resolve cleanly outside a project's own repository. Native gRPC features such as server streaming, client streaming, and bidirectional communication are largely unsupported, since most tools only display a single request-response pair. A protobuf-native workflow would instead offer a proto registry, schema-backed request building, native gRPC execution, and environment variables that work seamlessly across endpoints, metadata, and message fields.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Open Interpreter Lets Local AI Agents Execute Code, Not Just Generate It

Open Interpreter is an open-source tool that enables large language models to run code directly on a user's machine, moving beyond text generation into autonomous action. Developers building web apps and SaaS products can use it for tasks such as local data processing, automated code scaffolding, and on-device support diagnostics. Because it runs locally, it avoids vendor lock-in and reduces reliance on external API pricing for core agent logic. However, integrating it into production applications raises significant concerns around security, sandboxing, and user experience design. Challenges such as managing agent state across sessions and communicating clearly with users about what the agent can do will require careful architectural planning.

0 comments Read more at DEV Community