Local 30B models and a 750 tok/sec API tier

Hey, I haven’t been very active on Substack for quite some time. The reason is that I enjoyed building with AI agents more than creating long-form content.

I also realized that this space — agentic engineering — is moving incredibly fast. My day-to-day job looks completely different today than it did a year ago.

Part of why I started this Substack was to stay up to date myself and share the highlights. So I built a pipeline that pulls together the most important information, so I can consume it in five minutes. That’s what I’m sharing with you here.

Internally at Cresta I built something similar. It keeps the team current on what actually changed, and what it means for their day-to-day work.

This is for busy software engineers who work intensively with Codex, Claude Code, Cursor and other agentic harnesses. The curation comes from the sources I’ve read most over the past year — the ones that actually helped me become AI-native. You can read more about how items get chosen on the Start Here page.

I’ll keep working on this pipeline, to deliver the curation not just for you, but for myself.

16 links below, from 499 items across 47 sources this week.

Introducing Muse Glimmer
Meta’s return to open weights: a 30B Apache-2.0 vision model tuned for tool calling and multi-step agentic work, running locally in an 18GB quantization. A 32GB machine can now drive a coding-agent harness over your codebase with no hosted API, no per-token cost, and a permissive license.

Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
An OpenAI API service tier that serves GPT-5.6 Sol on Cerebras hardware at up to 750 output tokens per second. Agent loops that were unusable interactively at 50 tok/s become viable, and multi-step tool chains finish inside a single user’s attention span.

Introducing Grok 4.6
xAI’s latest model, tuned for multi-step agentic runs and matching GPT-5.6 Sol on the Artificial Analysis composite, shipping in Cursor and Grok Build. You can point long-running coding agents at it in Cursor today, with included usage doubled for the first week.

Linux desktop preview and agent imports
Codex desktop now installs on Ubuntu, Debian and Fedora, and imports instructions, skills, plugins and recent chats from Claude Code, Cowork or Cursor. Trying Codex against your existing setup costs one `/import` instead of a manual config rebuild, and sync keeps it current.

anthropics/claude-code v2.1.232
Claude Code release turning on forked subagents by default — they inherit the full conversation and prompt cache — and adding `@` mentions to message another live session. Fan-out no longer means re-sending context to cold subagents, and parallel sessions can now talk to each other by name.

Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things
Hands-on notes running the Apache 2 licensed 27B vision model locally via LM Studio at Q4_K_M, including its shipped default of xhigh reasoning effort. Set reasoning_effort to medium or low before wiring this into anything, or trivial prompts will burn minutes of local inference.

Using the GitHub Copilot SDK for Java
A client library for creating Copilot agent sessions, registering tools, and getting structured responses from server-side Java, with a Jakarta EE 11 sample app. You can embed agent loops in enterprise Java without Spring AI or Langchain4j, and point BYOK config at any OpenAI-compatible endpoint.

Cloud agents start 3x faster with builds
Cursor now rebuilds cloud agent environments hourly in the background, so sessions fork a live pre-warmed machine instead of cloning and installing from scratch. Cloud agents on a large repo start in seconds, and a commit that breaks your install script no longer blocks sessions — they fall back to the last good build.

React for Agents: Astro Creator Brings Hooks to his Meta-Harness, Flue
Flue 2, Fred Schott’s TypeScript agent framework, models an agent as a function that re-renders each turn, with 16 hooks like useTool() and useSubagent(). Tools, skills and subagents can be attached per-turn through composable hooks, replacing the state plumbing you hand-roll between model calls.

How many of your agent’s calls actually need a frontier model?
A 145-task benchmark of NVIDIA NeMo Switchyard routing agent turns between small and frontier models, with per-turn difficulty classification. Gives you a concrete tradeoff to price out: 74% less inference spend for six accuracy points, with only 7% of turns needing the big model.

TDD inside the agent loop - theater or actual value?
Birgitta Böckeler’s experiments testing whether instructing a coding agent to work test-first actually changes the code it produces, or just adds ceremony. You get evidence to decide whether “use TDD” belongs in your agent instructions, rather than cargo-culting it from human practice.

Your contributors are AI-first now. Is your project?
AutoGPT’s Nicholas Tindle on handling a PR queue largely written by agents: per-directory instruction files instead of a wiki agents never open. If your repo takes outside contributions, move machine-readable rules next to the code agents edit rather than into CONTRIBUTING.md.

Auto-research with codex: How I achieved a 232x Faster Kernel
A writeup of running Codex in an autonomous loop against a GPU kernel benchmark, letting it generate and measure variants until it hit a 232x speedup. Gives you a template for handing agents any task with a hard numeric target and a fast measurement loop, instead of reviewing each attempt yourself.

Maximizing the value of your Claude Code sessions
Anthropic’s own account of how to scope Claude Code sessions — when to reset context, when to plan first, and how to size a unit of work. Gives you a first-party baseline for session hygiene to compare against your own habits, particularly around when to clear context rather than push through.

What’s the best programming language for coding agents?
Dan Luu measures how many tokens equivalent programs consume across languages, arguing verbosity is a direct tax on an agent’s context budget. Gives you a cost basis for language choice in agent-heavy repos, where token-dense code means more context left for reasoning and fewer iterations.

Recovering Encrypted LLM Reasoning Traces
A hands-on reproduction of a published attack that reconstructs the plaintext of the encrypted reasoning blobs OpenAI and Anthropic APIs return to clients. Stop treating those opaque reasoning fields as confidential — anything the model thinks through is recoverable by whoever holds the blob.