A 27B model on a laptop matched Sonnet 4.5 on JetBrains' agent tests

Qwen3.6-27B at 4-bit, about 20 GB on disk, running on an M5 Mac with 64 GB of RAM, scored on par with Sonnet 4.5 on JetBrains’ private agent test set. That is the number underneath Junie Local, which JetBrains shipped yesterday: type /local inside Junie, the weights download, a local server starts, and the agent runs with nothing leaving the machine. Free, no account, works on a plane.

The engineering notes are more interesting than the launch. JetBrains argues the metric everyone chases for local models is the wrong one, because a coding agent spends most of its time in prefill reading files rather than generating tokens, which is why Junie Local starts at M5 rather than M4: the M5 Neural Accelerator’s 8-bit arithmetic instructions bought roughly 40% more prefill throughput, and they say the patch is going upstream to MLX-VLM. They picked Qwen3.6 over the newer 3.8 because 3.8 needs reasoning enabled to behave, and reasoning made tasks about four times slower. They ship with reasoning disabled entirely, having measured that it added little quality at two to three times the tokens. So the Sonnet-4.5-parity claim is a no-reasoning local model against cloud models with reasoning on. JetBrains is blunt that a 64 GB M5 is a steep requirement and says DGX Spark and RTX 5090 prototypes already work. The part worth borrowing is the argument about what changes when nothing is metered: multi-file refactors, dependency upgrades, and the test-coverage gap you have been ignoring for two quarters are exactly the work you stop asking for while watching a credit balance.

GitLab CEO Bill Staples published a long piece on the same economics from the other end, and it is the densest collection of production agent numbers I have seen in one place. Stripe merges more than a thousand pull requests a week produced entirely by its minions agents, which run in isolated environments and selectively draw from a suite of over three million tests. Amplitude cut pull-request cycle time from 5.2 hours to 44 minutes and frontend CI from about thirty minutes to three or four, tripled PR volume over six months, and watched reported monthly bugs fall from 715 to 319; non-engineers went from essentially zero PRs to roughly five percent. Spotify’s background agent Honk has more than 1,500 merged PRs in production. Staples’ framing is that the unit to measure is cost per accepted change, not cost per line, and that once generation collapses everything else in that term becomes proportionally larger. His sharpest line is operational rather than strategic: a thirty-minute CI pipeline defeats every model you point at it. The essay is also a GitLab platform argument, so discount the four-capability architecture accordingly, but the ninety-day checklist at the end (time your CI, write down merge criteria rather than approvals, make context portable) costs nothing to run against your own pipeline.

The mystery-model story from Saturday moved. A post on dejan.ai arguing that Ox-Alpha is a GLM model hit the Hacker News front page yesterday with 50 points and 27 comments, which is the first public attribution attempt since Business Insider wrote up the free model nobody could place. Treat it as an unconfirmed identification until Z.ai says something.

On the harness side, TrueFoundry open-sourced TrueForge, an enterprise agent harness whose headline claim is 30 to 75 percent cheaper task completion than Claude-managed agents. That comparison is the vendor’s own, on the vendor’s own tasks, and the interesting thing is not whether the range holds but that harness cost is now the axis vendors compete on publicly, a few days after DeepSeek open-sourced its own harness.

The best research item in the window is empirical rather than benchmark-shaped. Programming by Chat analyzes 74,998 developer messages across 11,579 chat sessions, 1,300 repositories, and 899 developers using Cursor and GitHub Copilot, drawn from chats committed to public repos during ordinary work rather than from a lab study. Three patterns come out: developers specify progressively rather than up front, refining outputs instead of writing complete task descriptions; they hand diagnosis, comprehension, and validation to the model rather than doing those themselves; and they manage the collaboration explicitly, externalizing plans into persistent files and negotiating autonomy through context injection and behavioral constraints. If you have ever wondered whether your AGENTS.md habit is idiosyncratic, this is the evidence that it is the emergent norm.

Meta published two hardware posts the same day. MTIA 300 is its first in-house training chip with NIC chiplets built into the accelerator and communication-offloading engines on die, aimed at ranking and recommendation training. The companion post covers MetaRoCE, a clean-sheet RDMA transport for AI-scale Ethernet. Both are the same bet: at frontier scale the interconnect, not the matrix multiply, is what stalls the GPUs.

Two smaller availability items worth knowing. OpenAI put GPT-5.6 into Kiro on a price-performance pitch, the same week GPT-5.6 Terra and Luna reached Bedrock in both AWS GovCloud regions. Google’s ADK added native evaluation for live voice agents, automated testing across multi-turn conversations, which is the least glamorous and most necessary piece of voice agent tooling.

And the pricing story practitioners actually argued about: Cursor Ultra’s revised 8/24 limits split into $500 of frontier-model usage plus $2,000 of Cursor’s own models. The split is the message. Vendors with in-house models are pricing them at four times the frontier allowance, which is a bet that most agent work does not need the expensive model and a nudge toward finding out.

What to watch: whether anyone reproduces JetBrains’ M5 prefill numbers on non-Apple hardware, and whether Z.ai confirms or denies the Ox-Alpha attribution.