Local AI tools dominate the week: Jevstiller, TalkScribe and a cloud retreat at Anthropic

Jevstiller is an open-source proxy that sits in front of TypeSafe's Jev. Its GitHub repository and a Register write-up both date to 30 September.

Both describe the same mechanism. Every request goes to Jev at first, and the answers train a small local model. Once enough samples exist, the familiar ones are answered on the user's own hardware. Jev is cheap, but not free. The Register puts Jev's price at $42 per billion input tokens and its latency at around 300 ms. Jevstiller's README says a local answer takes about 15 ms on CPU, roughly 20 times faster. The project's own documentation describes it as a cache: "A request is answered locally only if every question in it can be; otherwise the whole request goes to Jev with the caller's own key."

The numbers behind the claim are narrower than the headline. Banking77 is a set of 77 customer-support intents replayed against recorded Jev answers. On it, the default settings answered 61.7% of held-out messages at 99.40% agreement with Jev, against a stated 98% target. The README also reports accuracy of 78.8% for the local model against Jev's 78.5% on the dataset's labels. A live run on 25 September gave 70.7% at 99.45% agreement, using a setting the README says is no longer the default.

Local models move onto the laptop

The same week brought a cluster of tools built on one assumption: inference belongs on the machine in front of you. TalkScribe, a macOS dictation app published on 30 September, runs speech-to-text and a rewrite pass on Apple's Neural Engine, with no cloud transfer. It sells a lifetime licence for $14 and ships a Qwen3 1.7B model for polishing. Its site claims people type about 40 words per minute and speak 130 to 160. That is the pitch in one line.

Ollama added a decision-model endpoint in version 0.35, according to its library page. The endpoint serves Nimble, a 9B model from Bespoke Labs fine-tuned from Qwen3.5-9B. The page says Nimble answers in under 100 ms on a MacBook Pro with an M5 Max and takes up to 64 questions per request. The endpoint follows TypeSafe's Jev API, so the same shape of request works against a hosted Jev or a local file.

"We're open sourcing AA-AgentPerf-Local, which replays real agent trajectories on laptop & workstation hardware to test inference performance," Artificial Analysis wrote on 30 September.

That benchmark gives the local push a measuring stick. Artificial Analysis says its default workload is 8 recorded agentic tasks spanning 168 model turns, with context growing to about 56K tokens. It published initial results for NVIDIA DGX Spark at 128 GB, AMD Ryzen AI Halo at 128 GB, a MacBook Pro M5 Pro at 64 GB and an NVIDIA GeForce RTX 5090 at 32 GB. The models tested are Qwen3.5-9B, Qwen3.8-27B, Qwen3.6-35B-A3B and Ling 3.0 Flash at 4-bit quantization.

Two more projects fill out the pattern. OpenResearch, from alphaXiv, is a local-first workspace for research agents that keeps experiments in a local SQL database. It says it does not collect code or agent traces, though its release builds send opt-out analytics tied to a random installation ID. LocalThings, a Home Assistant integration, talks to Samsung appliances over the local network using CoAP over DTLS and the OCF specification. Its README says it never contacts Samsung's servers after setup.

The counter-example: Anthropic pulls local mode

Not every vendor is moving in that direction. Anthropic's help centre states that on 6 October 2026, new Cowork tasks on Pro, Max and Ultra plans began running in the cloud, and the "Only on your computer" option in Settings was removed. Users still set to local were moved to the cloud the next time they opened the desktop app, with no way to switch back. Tasks started before that date stay on the machine.

Put the two halves together and the week reads as a split, not a trend. The tools above are small, single-purpose and mostly free or cheap: a $14 licence, a container image, a GitHub repository. Anthropic's change affects paid plans and moves work the other way. Which side wins depends on workloads, and on whether the agreement numbers hold outside a 77-class intent dataset.