What happened
Kapitsa’s v0.2.36 release notes, published on GitHub on November 15, describe performance changes for 27B-model agent workloads and prompt prefill. The release identifies nine commits to the main branch since the preceding release. Its headline claim is that “27B agent turns on a Mac mini M6 are faster,” followed by three before-and-after measurements for different parts of an agent workflow.
First, the notes say a 45-row cached turn now takes 231 milliseconds, compared with 544 milliseconds previously. That measurement concerns an already cached interaction, rather than the first processing of a long prompt. Second, a compacted prompt takes 42 seconds instead of 72 seconds. Third, a whole agent session takes 140 seconds rather than 153 seconds. The release treats these as separate scenarios: a cached turn, a compacted-prompt operation, and an end-to-end session. In particular, the reported reduction for the full session is smaller than the gains reported for the two narrower measurements.
The release also changes the default hardware path for prompt prefill. The Neural Engine now participates in part of prefill by default on M1, M2, and M6-or-later systems. Kapitsa’s notes report that a 27B 16K-token prompt on an M1 Ultra takes 48 seconds rather than 61 seconds. They also report that a 4K prompt on an M6 takes 9.3 seconds rather than 16.1 seconds. The wording says the Neural Engine takes part of prefill; it does not claim that it executes the whole prefill operation. The announcement does not describe a user-facing switch, configuration command, or separate installation step for the new default.
Two additional changes cover long-context behavior and model packaging. Cold 128K prompts are no longer killed by macOS on an M2 Pro. The 27b alias now pulls the Nace Q4 pack, which the release notes describe as closer to the original model. The project published the update as release v0.2.36 in the nace-ai/kapitsa GitHub repository. The notes state neither a price change nor access restrictions, and they do not present the release as a new model or a new hosted service.
What the archive adds
New Runtime’s earlier coverage includes Kapitsa 0.2.15’s stability fixes and faster MoE prefill (October 24, 2026), Kapitsa 0.2.24’s Splash backend and Qwen3.5 MLX drafting support (October 31, 2026), and Kapitsa 0.2.25’s Clover model and C ABI SDK (October 31, 2026). This is a new release on a known project, with new performance claims rather than a restatement of those additions.
Trend signal
This release does not directly establish a broader agent-runtime pattern: its central claims concern local inference speed and hardware use, rather than security boundaries or the architecture of production harnesses. This does not yet fit a pattern New Runtime tracks.
Why it matters
For a team evaluating local models, end-to-end session time can matter more than a headline token rate. Kapitsa reports a smaller improvement for the complete session than for either the cached turn or compacted prompt. That distinction matters because an agent workflow includes more than prompt ingestion: tool calls, context changes, generation, and orchestration can all remain outside the part accelerated here.
The Neural Engine default makes hardware-specific testing more important, not less. A result on an M1 Ultra or M6 does not establish the same gain on every Apple Silicon machine, model pack, or prompt shape. The long-context fix is separately relevant for teams whose agents begin cold with large instruction sets, retrieved context, or persisted histories. A workload dominated by short cached turns may see a different operational effect from one that repeatedly sends fresh long prompts.
What happens next
- Kapitsa will publish further Apple Silicon performance or compatibility changes around prefill, long context, or model packs. (3 months, 65%). Watch for: release notes with comparable device-and-prompt timings.
- Teams using local 27B models will evaluate the default Neural Engine path against their own end-to-end agent latency. (6 months, 60%). Watch for: reproducible workload comparisons across Apple Silicon machines.
Limits and uncertainty
The release notes provide selected before-and-after timings but do not specify a full benchmark protocol, run counts, variance, thermal conditions, or whether all timings were measured under identical conditions. The figures are not independently verified here and should not be generalized beyond the named devices and prompt scenarios.
Kapitsa does not quantify memory use, output-generation speed, power consumption, or the effect on model sizes other than the cited 27B workload. It also does not state how much of each prefill operation moves to the Neural Engine, which operations remain elsewhere, or how the default affects workloads that do not benefit from that path. The macOS long-context claim names an M2 Pro but does not establish behavior for other machines.
Skill to build now
Create a small local-inference benchmark that mirrors one real agent task rather than measuring only output tokens. Run it in three modes: a warm cached turn, a cold long-context prompt, and a full session that includes the same compaction or context-management step your agent uses. Record wall-clock latency separately for each mode, then compare the complete session before and after upgrading. If your deployment runs on Apple Silicon, include at least one representative M1, M2, or M6 machine and retain the exact model pack and prompt lengths used.
Reader takeaway
Kapitsa 0.2.36 makes a specific default-path change for Apple Silicon and publishes concrete, scenario-bound timing claims. The practical question is not whether one benchmark is faster, but whether Neural Engine-assisted prefill reduces the full latency of the prompt shapes and agent sessions your team actually runs. Benchmark those sessions before revising local deployment plans.