Loading
Loading
Jev launched on a Monday. By Wednesday there were six open reproductions: a ModernBERT encoder, a diffusion model, three Qwen fine-tunes and a 40K byte-embedding scorer. They differ in backbone and agree on everything that matters — each one answers, forgets, and never learns whether it was right. We measured what happens when any of them sits behind a loop that does: a coin-flip decider reaches 92.5%, a 72.5% one reaches 96.1%, on a seventh of the calls. Plus the 0.6B embedder that lets a similar situation answer too.
TypeSafe AI raised a $40 million seed for Jev, a model that answers typed decisions and forgets every one of them. Independent tests put it at 72.5% accuracy with the weakest calibration of six models, graded at launch against other models' opinions, behind a waitlisted API. A decision system should learn from what happened, decline what it does not know, and run on hardware you own. Ours does all three: 97.5% on a tenth of the model calls, a repeat in 9 milliseconds, on one desktop.
Google has open-sourced AX, a Kubernetes-native orchestrator for agent workloads: isolated tasks, egress allowlists, suspend and resume, declarative apply. Those are the right primitives. awrun 0.2 ships the same four on one machine with no cluster, no Redis and no account — a run suspends in 2.3 ms, a resumed workflow replays the model calls it already paid for instead of making them again, a limit the runner cannot enforce fails the run rather than being ignored, and every lifecycle change is written to a hash-chained audit trail before it happens. Use AX if you already run Kubernetes at cluster scale. Use this if you want the same guarantees on the box under your desk today.
A bounded decision — pick A or B, score this, yes or no — does not need a frontier model on every call. Our decision door answers a judgement it has already been taught at 0 model calls and 0 dollars, reports a 60/40 coin as 0.60 instead of 0.99 once it has seen the flips, and went from 0.0% to 76.7% on a retrieval task after we fixed our own broken index and built a symbol index. It is calibrated on real journaled forks at ECE 0.091 — after we found that our own recorder had never written the probability column at all, which is also in here. The numbers, the table, and the command to reproduce them.
Meta announced Muse: a personal agent on a dedicated cloud VM with a shared browser, approvals, a credential store and a Mac app. Every line on that product page is a thing you can install from us today — awdk, awsh, awdesk and AitherOS Online — on hardware you own, on three operating systems, with the audit trail in your hands. Here is the page, line by line, and the four commands that replace it.
Bonsai 2 27B is a ternary model — every weight is -1, 0 or +1 — and it ships with a vision tower. On our 5090 it answers image questions in 1.4 s, sustains 54.6 tok/s against 9.5 for the 12B multimodal we ran before it, reads a bar chart correctly where the 12B did not, solves an ARC-style grid in 2.2 s, and takes sixteen video frames in one prompt where the 12B caps at four. Use it for everything except literal OCR of small text, where the 12B is still the better transcriber. Here are both models on the same prompts, the same images, and the command to rerun them.
Everyone else ships a function that calls a model every time. We built the loop: the same brain goes from 72.0% to 97.5% accuracy on a tenth of the model calls, a repeated decision drops from 35 seconds to 9 milliseconds with no model involved, and a judge scores 100% at zero calls. One desktop, seven benchmarks, one command to reproduce all of it.
We benchmarked PrismML's Ternary Bonsai 2 27B against Bonsai 1 on a single RTX 5090. It scores 0.9165 to 0.7821 on AitherBench, answers every item, and does it in 44% of the tokens. The 5.95 GB ternary quant matches the 7.2 GB one exactly. Run Bonsai 2, pick PTQ1_0.
AitherTunnel went live on 2026-03-09: per-node tokens, zero-disk WireGuard peers, agents that join a private mesh with their own identity, a Mac-mini-under-the-desk reachable from a phone. On 2026-04-14 Cloudflare announced Cloudflare Mesh with the same shape, feature for feature. Here is the timeline, what ours does that theirs does not, and how to run it today.
Five months ago we published an 8-layer maturity ladder for AI engineering. Running it in production, with agents doing most of the work, gave us the upgrade: two questions asked at every layer, a short rulebook for each, and one number to grade yourself on.
A 475 GiB model, 96,085 tensors, and an architecture that reuses attention state across layer groups. What unlocked it was reading the inference code DeepSeek shipped next to the weights — and then two segfaults taught us the part we still had wrong.
We distilled a code-search student, then pointed it at a customer's ordinary prose. It ranked better than Qwen3-Embedding and nomic — and better than the fallback model that was actually serving those documents. Nine queries in the table were real user questions.
The same conversion that runs our code-search embedder in a browser tab now serves it in production too: the Q8_0 GGUF, through llama.cpp instead of a full-precision engine. Measured side by side before the switch: 0.9997 agreement with the original weights, identical search geometry, 7 ms per embedding instead of 31, healthy in 9 seconds instead of 90, and about a gigabyte of GPU handed back. Here is how one artifact ended up serving three places, and the two checks that made the switch safe.
The 0.6B student that out-retrieved its 7.85B teacher is now a 396 MB file that the browser downloads once and runs itself. Cold start to first vector in under 19 seconds, about a second per document after that, 0.99 agreement with the full-precision model, and your question never leaves the tab. Here are the numbers, what it plugs into, and the two metadata details that make a GGUF embed correctly anywhere llama.cpp runs.
We distilled a 7.85B-parameter embedding model into a 0.6B student for searching our own codebase. The student retrieves the right directory on the first try 80% of the time against the teacher's 66%, fits in 1.06 GB as int8 with 0.999 fidelity, and the whole run — capture, train, quantize, eval — was one command and about three dollars of rented GPU. Here are the numbers, the recipe, and the honest reason a small model can out-retrieve a large one.
awrun can see that CI is saturated. It cannot buy a server, and that is on purpose. Here is the split we shipped — the queue decides, the host acts — and the measured reason we found it: an autoscaler that had been running every five minutes for weeks and had never once added a runner.
We measured every way to serve DeepSeek-V4-Flash — 284B parameters, 90.9 GB quantized — on hardware we already own. The winner is not the one the architecture diagrams assumed. 24.2 tok/s median, single box, with a 0.6B drafter doing the proposing. Here's the honest A/B behind it, the two-box split that's still significant, and the recipe we made public.
We measured the real-world performance of in-browser session memory — embeddings, long conversations, and transparent resource usage. The results are clear: faster, more efficient, and more controllable than you ever imagined.
Technical deep dives and build updates. No spam, no fluff.
Subscribers get first priority when invites open.