OpenRouter

Technical Primer on Kimi K3 and KDA Inference Kernels

This SemiAnalysis technical primer explains the core techniques behind the Kimi K3 model architecture, focusing on Kimi Delta Attention (KDA), its ancestry (linear attention, DeltaNet, Gated DeltaNet), and system-level implementations such as FlashKDA. The article describes FlashKDA’s kernel design (K1/K2), arithmetic and memory complexity (O(T*D^2) compute at chunk level, prefill linear, decode constant), interactions between KDA and full-attention MLA in hybrid architectures, attention-residual and block-attention-residual designs, LatentMoE communication trade-offs, and load-balancing via Quantile Balancing. The piece also reports benchmarking activity (InferenceX), day‑0 bringup notes (NVIDIA/AMD recipes on vLLM), and observed OpenRouter provider pricing floors as of July 30.

One API key to compare LLM token costs

The author recommends placing a thin request router in front of an application to use a single API key while comparing token costs across OpenAI, Anthropic (Claude) and Google's Gemini. Token sticker rates are often misleading because input tokens (retrieved context, system prompts) can dominate costs and retries or eval harnesses can dramatically raise spend. The article describes reading live model catalogs (example: Infrai) and counting tokens via a token-counting endpoint before sending requests, pricing calls using per-input and per-output per-million-token fields, and routing by cost while reserving direct vendor SDK calls for vendor-specific features (e.g., Anthropic prompt caching, Gemini large context windows). Practical implementation tips include honoring Retry-After, avoiding hardcoded rates, logging estimated costs, and refusing expensive eval runs.

Use One API Key for Multiple Text-to-Image Models

A developer guide arguing that if text-to-image generation is one feature among many, teams should put image generation behind a unified API so a single credential can reach multiple vendors; pin the selected model id in configuration and treat provider selection as a deploy-time concern. The author contrasts when to use an aggregator (reduced integration overhead, single wire format, unified billing/metrics) versus integrating a vendor directly (faster access to new parameters, fine-grained GPU controls, compliance/region control). The post shares operational lessons (idempotency keys, read-back verification, honoring Retry-After) and references vendors including Infrai, OpenAI, Google Gemini, Anthropic/Claude, Replicate, OpenRouter, and Amazon Bedrock.