---
type: "post"
id: "kapitsa-0-2-34-adds-an-engine-page-196b65"
slug: "kapitsa-0-2-34-adds-an-engine-page-196b65"
title: "Kapitsa 0.2.34 adds an /engine page and ships KQuant 4-bit packs"
description: "Kapitsa 0.2.34 adds an /engine page for inspecting and controlling a running server, alongside KQuant 4-bit model packs and Apple Silicon performance claims."
retrieval_nugget: "Kapitsa 0.2.34 adds an /engine page for inspecting and controlling a running server, alongside KQuant 4-bit model packs and Apple Silicon performance claims. For teams running models locally, the release combines more operational visibility with a new quantization option, but its speed and quality comparisons remain the project’s own claims."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["inference"]
entities: ["Kapitsa","Qwen3.6 35B-A3B","Qwen3.5 4B","Qwen3.5 9B","Qwen3.8 27B"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Kapitsa 0.2.34 adds an /engine status page and ships KQuant 4-bit model packs."
prior_refs: ["https://newruntime.com/posts/kapitsa-0-2-24-starts-a-splash-engine-7843ad/","https://newruntime.com/posts/kapitsa-0-2-22-runs-flash-next-from-9160f3/","https://newruntime.com/posts/kapitsa-0-2-25-adds-clover-35b-a3b-58bbd7/"]
source_url: "https://github.com/nace-ai/kapitsa/releases/tag/v0.2.34"
source_urls: ["https://github.com/nace-ai/kapitsa/releases/tag/v0.2.34"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:kapitsa-0-2-34-adds-an-engine-page-196b65"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/kapitsa-0-2-34-adds-an-engine-page-196b65/","markdown":"https://newruntime.com/posts/kapitsa-0-2-34-adds-an-engine-page-196b65.md","json":"https://newruntime.com/posts/kapitsa-0-2-34-adds-an-engine-page-196b65.json"}
markdown_url: "https://newruntime.com/posts/kapitsa-0-2-34-adds-an-engine-page-196b65.md"
---

# Kapitsa 0.2.34 adds an /engine page and ships KQuant 4-bit packs

## Retrieval answer

Kapitsa 0.2.34 adds an /engine page for inspecting and controlling a running server, alongside KQuant 4-bit model packs and Apple Silicon performance claims. For teams running models locally, the release combines more operational visibility with a new quantization option, but its speed and quality comparisons remain the project’s own claims.

## Summary

Kapitsa 0.2.34 adds an `/engine` page for inspecting and controlling a running server, alongside KQuant 4-bit model packs and Apple Silicon performance claims. For teams running models locally, the release combines more operational visibility with a new quantization option, but its speed and quality comparisons remain the project’s own claims.

## What happened

Kapitsa’s GitHub release notes for version 0.2.34 say the new `/engine` page shows memory, slots, drafts and requests. From that page, a user can stop the server or switch models. The notes present this as a way to see what Kapitsa is doing while it runs and to take two direct actions without leaving that view.

The status view is notable because it joins observation and intervention in the same surface. Memory is one of the displayed signals, alongside the slots used by the engine, drafts and active requests. The two actions named in the release are deliberately bounded: stop the service, or change the selected model. The notes do not describe alerting, access controls, multi-user operation, remote administration, or additional actions from the page, so those should not be inferred from the announcement.

The release adds KQuant 4-bit packs for Qwen3.6 35B-A3B and Qwen3.5 4B and 9B. The notes say KQuant stays closer to the original model than the packs it replaces at the same size. That statement describes a relative quality goal at a fixed pack size, rather than promising a particular benchmark score or improvement for every task. The release does not name the older pack formats being replaced in this comparison.

For Qwen3.8 27B, KQuant is opt-in: the specified command is `kapitsa serve --model 27b:kquant`. The release also says Qwen3.8 27B writes about 8% faster on an M5 Pro when using Kapitsa’s own drafter. This is a separate claim from the KQuant quality statement: it concerns writing speed with a particular model, machine class and drafting arrangement.

A further performance claim concerns prompt reading. Kapitsa says prompts are read up to 1.9 times faster on M5 Macs for 3-bit, 8-bit and GGUF 4-bit models. The stated test configuration is 4K tokens with Qwen3.5 4B. The “up to” figure and this specific configuration matter: it is not a general claim that every model, prompt length, quantization format or workload will see that speedup.

The notes also credit smarter SSD cache eviction. When a conversation is replaced, it now leaves the SSD prefix cache first, so live chats keep their cache. A further change makes IQ2_S CPU decoding about a third faster per block under load, with the same output, according to the release. The notes identify contributors to these two changes as yonaskolb and chainhelen. GitHub marks the release as published on 10 November 2026, and notes 99 commits to the main branch since the prior release.

## What the archive adds

New Runtime’s earlier coverage includes [Kapitsa 0.2.15’s Apple Silicon inference updates](https://newruntime.com/posts/kapitsa-v0-2-15-an-open-source-apple-e8f061/) (24 October 2026), [0.2.22’s EXL3 support and reported M5 Max benchmark](https://newruntime.com/posts/kapitsa-0-2-22-runs-flash-next-from-9160f3/) (28 October 2026), and [0.2.24’s Splash backend and Qwen3.5 MLX drafting support](https://newruntime.com/posts/kapitsa-0-2-24-starts-a-splash-engine-7843ad/) (31 October 2026). This is a new release in an ongoing sequence of Kapitsa capabilities, not a restatement of those earlier updates.

## Trend signal

This update supports [Harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/) in a limited, operational sense. The `/engine` page exposes runtime state and controls separately from the selected model, while the release also expands the set of model-pack choices. That separation is useful evidence that the operating layer can matter alongside model choice. The release does not show a broader production harness or demonstrate that this boundary is durable across model changes, so the connection is suggestive rather than conclusive.

## Why it matters

Local inference makes teams responsible for both model selection and the machinery around it. A status page that surfaces memory, requests and drafts can make a running process easier to inspect, while stop and switch controls reduce the distance between noticing a problem and responding to it. For a developer working at a terminal or managing a local machine, this can be a more direct operating surface than treating the inference server as an opaque background process.

KQuant also gives users another size-constrained choice. If the release’s quality claim holds for a team’s workload, it could offer a way to retain more of the original model’s behavior without moving to a larger pack. The opt-in command for Qwen3.8 27B makes the selection explicit, rather than silently changing the model pack behind an existing invocation.

The practical implication is to treat these features as options to validate, not automatic upgrades. Compare the packs on representative prompts, measure latency on the hardware and workload that matter, and verify that the new controls expose the state operators actually need. The release’s benchmark setup is narrow enough that teams should not assume the quoted gains transfer unchanged.

## Limits and uncertainty

The release notes do not provide benchmark methodology, sample counts, variance, or independent measurements for the 1.9x prompt-reading and roughly 8% writing claims. They also do not quantify how “closer to the original model” was evaluated for KQuant, or specify the affected pack versions beyond the named models. The SSD cache and IQ2_S changes are described without a full test setup.

These are project release claims. Performance and quality on other Macs, model variants, prompt lengths, quantization choices or workloads have not been independently established here. The notes also do not state whether the `/engine` controls require a particular deployment configuration, how access to them is governed, or how a model switch affects in-flight requests. Teams with shared or production-like local environments should establish those operational details before relying on the interface.

## Skill to build now

Create a small local-inference acceptance check before changing quantization or drafting settings. Run a representative prompt set through the current pack and KQuant at the same model size, then record output quality, prompt-read latency and generation speed separately. If a team uses Qwen3.8 27B, make the KQuant choice explicit in the launch command and test it with the project’s own drafter rather than assuming the reported M5 Pro result applies to another setup.

Use the `/engine` page during that test to observe memory, slots, drafts and requests while the run is active. Stop or switch models only in a controlled test session, then decide whether the page provides enough operational context for the people who will own the local runtime.

## Reader takeaway

Kapitsa 0.2.34 adds useful local-runtime visibility and KQuant packs, but the most consequential numbers are tied to particular models and Apple Silicon configurations. The release is most useful as a testable operating and quantization update: inspect the new runtime view, compare KQuant against the pack already in use, and make deployment decisions from measurements on the workload that matters.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
