Inflect-Micro-v2 runs a full offline TTS in under 10M parameters
Apache-2.0, 6x faster than real time on a plain CPU, no cloud. Plus: a 28.9M LLM on an $8 chip and Anthropic's new rulebook for prompting Claude 5.
Copy markdown
Full speech synthesis, under 10M params
Inflect-Micro-v2 does complete text-to-waveform in 9,356,513 parameters — a 37.5MB FP32 file that hits 6.28x faster-than-real-time on a plain 8-vCPU CPU, no GPU or cloud. Output is 24kHz mono with a fixed English voice and deterministic seeds, so the same text and seed always give byte-identical audio.
Apache-2.0, so you can ship it in anything
The weights are Apache-2.0 with a PyTorch-free ONNX Runtime path, so you can drop offline voice into a desktop app, game, or kiosk with zero per-call API cost or network round-trip. It's English-only for now, and there's an even smaller Nano-v2 if <10M is still too heavy for your target.
Meanwhile: a 28.9M LLM runs on an $8 chip
A developer squeezed a 28.9M-parameter model onto an ESP32-S3 microcontroller (~$8), generating ~9.5 tokens/sec fully offline. The trick — per-layer embeddings streamed from flash, borrowed from Gemma — keeps the 14.9MB 4-bit model out of the chip's 512KB of SRAM. It only writes TinyStories, not code, but the recipe is a real on-device weekend project.
Anthropic rewrote how you prompt Claude 5
Anthropic's new context-engineering guide reports cutting 80%+ of Claude Code's system prompt for Opus 5 with no eval loss. The new rules favor judgment over hard constraints, progressive disclosure through Skills, deferred tool loading, and auto-memory over a hand-edited CLAUDE.md. Run /doctor in Claude Code to audit and shrink your own context.