Ephemeris.
Four ways the machinery got sturdier — models that serve themselves, run in your browser, prove their own work, and shrug off dead hardware.
Write the model once, serve it fast.
For years, making a model quick in vLLM meant keeping a second, hand-tuned copy of it alongside the Transformers one. The new modeling backend closes that gap: pass --model-impl transformers and dynamic layer fusion gives any compatible model native vLLM throughput. Across Qwen3 at 4B, 32B and a 235B MoE, the single implementation meets or beats the hand-written version. One file to keep in sync instead of two.
Ship the model to the browser, not the server.
LiteRT.js drops Google's native inference runtime into the browser through WebAssembly and WebGPU, so a model runs entirely on the user's machine — no server round-trip, no upload of their data. The demo runs real-time object detection with Ultralytics YOLO. If you've been paying to host inference that could live on the client, this is the week to move it.
- 3× faster than other web runtimes, CPU and GPU
- WebGPU acceleration, WebNN for NPU hardware
- Local inference — private by default, no server bill
A machine that shows its proof.
Leanstral 1.5 is an open model built for one job: writing machine-checked proofs in Lean 4. At 119B total but only 6B active parameters, it saturates miniF2F and clears most of PutnamBench. The practical payoff is verification that catches real defects — the team used it to surface five previously unknown bugs across 57 open-source repositories, with weights and a free API within reach of an ordinary project.
- miniF2F
- saturated
- PutnamBench
- 587 / 672
- FATE-H
- 87%
- Bugs found
- 5 across 57 repos
We killed a TPU mid-run. It recovered in two minutes.
Train on enough chips for long enough and hardware will die on you; the usual answer is to tear the whole job down and reschedule. MaxText's elastic training turns a dead TPU into a caught Python exception — PathwayS flags it, an elastic_retry decorator reloads the latest Orbax checkpoint onto surviving chips, and the run continues on the slice that lived.
# 48 chips · 3 slices · Qwen3-0.6B · ~$30 / 30 min step 512 loss 2.41 throughput ok ✗ slice-2 lost — hardware fault raised as exception elastic_retry: reload Orbax checkpoint → surviving chips elastic_retry: rescheduling only the dead slice ✓ resumed in 1m52s · ~88 steps replayed step 600 loss 2.39 run intact
That's Sunday.
A quieter day on the feeds, so four picks over five — models pulling closer to the metal: native serving, browser inference, machine-checked proofs, and training that survives dead silicon.
Sources today
- Hugging Face
- Google Developers Blog
- Mistral AI
Rubric
- Adopt this week
- Dev tools & agentic coding
- Research w/ a practical kernel
- No reprints · 30-day cap
Issue
- No. 083
- 12 July 2026
- Zürich · 08:00