The story around Jev, the judgment-specialized model, moved past the rush of hands-on write-ups and into a new stage today: the arrival of an open implementation. In the local LLM scene, too, home-operations knowledge keeps steadily piling up — from builds that revive junk hardware to run 27B-class models to operating notes from expensive ARM machines.
Open Jev Implementation 'openjev' Arrives — and Plays Doom from Raw Pixels
Jev (TypeSafe AI) made big waves this week as a model that generates no text at all, returning only typed judgments and probabilities. Following the sudden flood of hands-on reports, an open implementation based on Qwen3.5 — **openjev** — has now been published on Hugging Face.
openjev is Qwen3.5 trained as a cross-encoder for NLI (natural language inference). It reads a pair of a premise and a hypothesis and returns one of three values: entailment, contradiction, or neutral. By design, this single primitive handles everything — reranking answers, grading against reference solutions, content guarding, even controlling real-time games — with no additional per-task training required.
The new 4B v2 checkpoint also supports images: it **plays Doom by reading raw pixels directly**, averaging 10.4 kills per episode (about double v1's figure, versus 1 kill for random play), and in Minecraft it reportedly completed the process of crafting an iron pickaxe from scratch in about 22 decisions. Benchmark numbers were published as well — 0.42→0.63 on the adversarial NLI dataset ANLI r3, and 0.52→0.84 on image claim verification — and a version with a 35B-A3B MoE backbone is included too.
On top of that, the browser-use team released **jev-ultrafast**, which speeds agents up by reducing browser actions to multiple-choice questions. A demo that completed the entire flow — searching Zurich→London flights on Google Flights and displaying the results list — in **7.1 seconds** has drawn attention, and code-reading explainers have already appeared in Japan. Validation of just how much practical speed this approach of never spending generation on judgments can deliver looks set to spread quickly.
Sam Altman to Brief the UN Security Council Next Week
According to Reuters, OpenAI's Sam Altman is scheduled to brief the UN Security Council next week. It is highly unusual for the head of an AI company to explain things directly before the Council, and amid international AI governance debates involving the US, China, and others, what he says — and how far he goes — will draw close attention.
The 'Traps' You Learn After Running Local LLMs on a GB10 for Months
A real-world account of running local LLMs almost every day for several months on a Lenovo ThinkStation PGX with an NVIDIA GB10 (128GB unified memory) was posted on Zenn. It covers an actual workload — an always-on agent, nightly batch jobs, and occasional OCR with a VLM — and sticks to 'a record of the pitfalls actually stepped in during operations, leaving the spec numbers to others.' The traps are ordered by how hard the damage was to notice, closing with 'a small experiment that tumbled in while researching the traps.' For anyone considering a unified-memory machine purchase, it looks like useful decision material.
Qwen 3.8 27B at 30 tok/s on Two CMP Mining Cards Totaling $250
On Reddit's LocalLLaMA, a build is drawing attention: a 10-year-old i7-4790K CPU fitted with two cards in total — a modified CMP50hx (20GB) and an unmodified 10GB version, **$250 combined** — running Qwen 3.8 27B at roughly 30–35 tokens/second (with MTP enabled).
The key is the arrival of a new driver patch that relaxes the PCIe 1.1 bandwidth limitation. Patching took some fiddling, but the builder reports eventually getting PCIe 2.0 operation working on the Z97 chipset as well (though the driver crashes on sleep resume, so the system runs with sleep disabled). After limiting power to 180W to give the PSU headroom, they also created a Q6K-quantized 130k-context preset and a 256k-context preset (10–14 tok/s). The pains unique to reviving junk hardware are candidly documented too, such as model loads from a SATA SSD taking one to two minutes.
GPU Rental Marketplace 'Clore.AI' — Unfazed Even When Used for Crime
LocalLLaMA also saw a warning post about the GPU rental marketplace Clore.AI. The poster tried to monetize by renting out their home rig, then discovered that a renter was using their connection to run vulnerability scans and distribute malware to a Colombian gambling site. They claim that after reporting it to the platform with archived evidence, the platform not only refused to act but blocked them from the support chat.
Tangled up with a terms-of-service clause forbidding inspection of renters' environments 'unless required by law,' the post throws the legal and security risks of lending out your home connection into sharp relief. Several other services let you earn money by renting out idle GPUs, but their response policies when abuse is discovered should be part of the decision too.
iOS 27's Foundation Models Now Let You 'Plug In Any LLM'
The developer of Sumibi, a Japanese input system, published an article examining Apple's Foundation Models framework, which changed significantly in iOS 27. The API, previously exclusive to Apple's on-device models, reportedly now takes a form where you can **plug any LLM in behind the same API**.
The iPhone version of Sumibi currently uses a BYOK approach, connecting directly to a user-configured OpenAI-compatible API. From the perspective that if Apple's models become usable, the app might run without any API key setup, the developer lays out a plan to test incrementally 'how much can be replaced' rather than migrating fully to on-device. Whether Apple will open its on-device AI as a foundation swappable with external models is a change with uneasy implications for future app development.
---
While the Jev boom broadens its reach in the form of an open-source implementation, local LLM self-build and operations knowledge piled up today at both extremes — 'a 10-year-old CPU + junk GPUs' and 'the latest unified-memory machine.' How Jev will change agent implementations as 'faster judgment' is the question, and validation results for openjev and jev-ultrafast look set to keep coming.