The week's clearest infrastructure signals came from runtime optimization: audio consolidation, in-browser inference, and speculative decoding convergence across a five-day window. A design-system format for agentic codegen, and a quantized model upgrade that added half again as much throughput without touching output quality, complete the week's five signals.
audio.cpp is the llama.cpp of audio, shipping this week as a single C++/ggml binary consolidating twelve TTS models (Qwen3-TTS, PocketTTS, VeVo2, and more) with reported 5x throughput gains on CUDA over equivalent Python stacks. The headline is the speed improvement. The more durable gain is operational: one dependency surface, one interface, twelve supported models. Anyone managing voice agent pipelines today is absorbing per-model Python runtime overhead and fragmentation across separate stacks. audio.cpp collapses that surface. The open question for any voice agent operator is whether the twelve supported models cover the quality and latency profile their pipeline needs.
On the browser side, Liquid AI's 230M parameter model is running at 1,400 tokens per second in-browser using hand-tuned WebGPU kernels rather than off-the-shelf WebGPU bindings, which accounts for the throughput gap over prior in-browser benchmarks. Twelve months ago that performance level in a browser runtime wasn't a realistic target for any capable model. For operators building zero-backend or privacy-constrained inference, LFM2.5 is the working reference implementation of what's currently achievable.
Speculative decoding was the week's convergence story. Three independent advances in five days: Eagle3 earlier this week, NVIDIA's parallel decoding architecture on the Nemotron backbone, and JetSpec's parallel tree drafting claiming 9.64x lossless speedup and throughput above 1,000 tokens per second by exploring multiple draft token branches simultaneously. The approaches differ technically: draft models, diffusion-based parallel decoding, tree-branch exploration. The operational outcome across all three is the same: major throughput gains with no quality tradeoff. Three independent convergences on the same technique in five days moves speculative decoding past the optimization-experiment category. The technique is no longer appearing in papers only; NVIDIA's Nemotron-TwoTower ships it as architecture, not a research claim.
The throughput gains compound only when the agents producing output remain stable across context resets.
Google Labs shipped DESIGN.md, a format spec for giving coding agents persistent, structured design context: typography, color tokens, component rules, brand voice. The conceptual model is CLAUDE.md applied to visual identity rather than project context. As agentic codegen pushes further into UI generation, design drift across context resets has become a real operational friction point. A repo-level spec that survives those resets addresses the problem at the right layer. The obvious use case is any repo where coding agents are generating frontend components alongside backend logic.
Gemma4-QAT added Multi-Token Prediction this week, delivering a 35% throughput boost on the 26B-A4B and 53% on the 31B-QAT with no reported quality tradeoff. The model family has been in this feed three consecutive days; MTP is the concrete new development that separates this week's entry from the prior two days of coverage.
That 1,400 tokens per second in a browser tab, with no server in the loop, is the week's most concrete number to carry forward. A year ago it wasn't a benchmark anyone was running.