Gigatoken: Fast Tokenization Without Changing the Tokens
A source-based look at Gigatoken's claimed tokenizer speedup, its SIMD and cache design, the benchmark conditions, and its current limits.
Back to Blog
CHANNEL / ACTIVE
#llm
Liquid Foundation Models: A Different Architecture for Edge AI
A source-based look at Liquid AI's continuous-time models, their proposed hybrid design, edge deployments, and the limits of the transformer comparison.
Agent Harness Architecture: Runtime Controls Around an AI Agent
A practical model of agent harnesses: tool execution, context assembly, sandboxing, state, and bounded error recovery.
Kimi K3: An Open-Weight MoE Model at Scale
Kimi K3 is Moonshot AI's 2.8-trillion-parameter open-weight Mixture-of-Experts model. A practical look at its architecture, the vendor's evaluations, API pricing, deployment limits, and the security responsibilities that come with open weights.
LLM model routing: choosing a model for each request
Model routing selects among models according to task, quality, budget, latency, availability, and data-residency requirements. A practical guide to routing policies, evaluation, cache locality, and their tradeoffs.
Tool-format translation between LLM providers
Tool calling looks similar across providers, but message shapes, streaming, reasoning state, and caching differ. A practical guide to what a translation proxy must preserve and when a higher-level integration is a better fit.
The harness matters as much as the model
An agent's model is only one part of its behavior. Prompts, tools, context, recovery, evaluation, and memory form the harness around it. What current research says, and how to measure its effect on quality and token use.
Using GonkaGate as an OpenAI-compatible model provider
A practical guide to connecting an OpenAI-compatible GonkaGate endpoint to command-line agents. How to create and protect an API key, test the endpoint, configure a custom provider, and verify models and costs before a workload.
OpenClaude for multi-provider coding work
How OpenClaude keeps cloud and local model providers in one terminal workflow, where its provider profiles help, and what to verify before moving a coding workflow.
Clother: separate Claude Code launchers for separate providers
How Clother keeps provider settings scoped to a Claude Code launch, how to inspect a profile before use, and where wrappers stop helping.
Reducing token use in Claude Code
A practical, measurement-first guide to lower context cost in Claude Code: smaller sessions, focused delegation, retrieval, compact tool output, caching, and API batching.