Kev's open Jev alternative now trains on Qwen3.5 for about $95
The decision-model category Jev created a week ago now has a full open implementation, a cost figure that undercuts every "frontier model" story this month — and a benchmark table that still puts Jev on top.
Jared Palmer ported Kev, his open decision-model family, onto Qwen3.5 bases and published the 0.8B, 4B and 9B checkpoints under Apache 2.0 — and the experiment log shows the whole Qwen3.5 effort cost about $95 of rented H100 time, plus three cents of Jev API calls. Kev is not a chatbot: each checkpoint is a rank-16 LoRA adapter and a small readout head bolted onto a Qwen base whose weights stay frozen, and it returns calibrated probabilities for typed questions instead of generating prose. The intended jobs are the small judgments buried inside software — routing a support ticket, picking an escalation path, scoring how frustrated a customer is.
The numbers are the interesting part, and Palmer publishes the unflattering ones. On the repo's locked, new-source test set Kev-9B scored 0.837 accuracy, up from 0.780 for the previous Qwen3-based Kev-8B; Kev-4B hit 0.832 and the smallest model 0.668. Hosted Jev still leads at 0.857 against Kev-9B's 0.812 on the development comparison — a gap Palmer flags as uncontrolled, since Jev's training data has never been disclosed. His pre-registered development criteria also went unmet even as the locked test favored the new checkpoints, and both results are in the repo rather than quietly dropped.
Two honest caveats come with the release. Qwen3.5's recurrent layers broke the block-attention trick the earlier version used, so each question now runs in its own row with the shared state computed once and its cache reused — which costs speed: on an M5 a five-question request takes 779 milliseconds on the Qwen3.5 Kev-4B against 174 milliseconds on its predecessor, and Kev-9B needs about two seconds. Palmer tells Mac users to stay on the older Qwen3 checkpoints until an MLX backend lands. The bigger context is that this is the third independent rebuild of the Jev idea in a week — we covered APUS reverse-engineering Jev into a fully offline browser agent yesterday, and the original Jev idea sitting unnoticed for a year before that. The category is being commoditized faster than the company that named it can monetize it.
Moonshot's Kimi K3 is now generally available on Amazon Bedrock, the first time the 2.8-trillion-parameter open-weight model is served from a major US cloud. AWS says the model keeps native vision and its one-million-token context window on Bedrock, with explicit prompt caching to cut latency and input cost on long coding and document workflows. The commercial terms are the part worth reading closely: AWS states prompts and completions are processed inside the AWS data boundary, are not shared with the model provider, and are not used to train the model — with zero operator access during inference.
That directly answers the sticking point from Moonshot's cloud talks. In August, Moonshot wanted a cut when US clouds sell its Kimi K3 model and, per Reuters, was asking for data access to audit token usage on its own model. Bedrock's launch terms grant the revenue channel without the telemetry. Whether Moonshot's audit ask survives on Azure and Google Cloud is now the question.
Apple's A20 Pro is the first iPhone silicon demonstrated running a 27-billion-parameter model on-device at roughly double the speed of last year's flagship. Developer Adrien Grondin ran Bonsai 27B on an iPhone 18 Pro and reported the phone beat his expectations for local inference, attributing the gain to the A20 Pro's dual 16-core Neural Engine and the 12 GB of faster LPDDR5X memory in the 18 Pro. It is a demo, not a benchmark — no tokens-per-second figure was published — but it is the first public evidence that the memory bandwidth Apple added is being spent on real models rather than assistant workloads.
The ceiling is still memory, not compute. Bonsai 2, the newer 27B build we covered in Bonsai 2 27B puts a 27B model on a laptop in 5.9 GB, is too large to fit locally even at 2-bit quantization on a 12 GB iPhone, so the demo runs the older generation. On-device 27B remains a proof of direction until Apple ships a bigger memory configuration.
What to watch: whether anyone outside the Kev repo reproduces its locked-test scores, and whether Moonshot's audit clause shows up in the next cloud deal.
Jev is still ahead on Kev's own benchmark table — does an open clone at 95% of the accuracy change who you'd ship? Tell us in the comments.
Sources: jaredpalmer/kev (GitHub)
