---
type: "post"
id: "kapitsa-0-2-23-lands-an-ordering-fix-75ad97"
slug: "kapitsa-0-2-23-lands-an-ordering-fix-75ad97"
title: "Kapitsa 0.2.23 lands an ordering fix aimed at Nemotron-H garbage output"
description: "Kapitsa 0.2.23 changes the Flint fused-decode path so merges arrive in the order expected by Nemotron-H."
retrieval_nugget: "Kapitsa 0.2.23 changes the Flint fused-decode path so merges arrive in the order expected by Nemotron-H."
published_at: "2026-10-28"
updated_at: "2026-10-28"
record_date: "2026-10-28"
date_kind: "published_at"
topics: ["github"]
entities: ["Kapitsa","Flint","Nemotron-H","Apple M5 Max"]
related_patterns: ["harness-architecture-outlives-model-choice"]
novelty_label: "new_fact"
what_changed: "Kapitsa 0.2.23 is a later release adding an ordering fix aimed at Nemotron-H garbage output, beyond v0.2.15’s stability and MoE-prefill changes."
prior_refs: ["memory:b1983391-b1a5-8fb5-ac96-81bf579d72ec","memory:6b651b8c-7011-8ab4-aad8-a0387c24f59c"]
source_url: "https://github.com/nace-ai/kapitsa/releases/tag/v0.2.23"
source_urls: ["https://github.com/nace-ai/kapitsa/releases/tag/v0.2.23"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:kapitsa-0-2-23-lands-an-ordering-fix-75ad97"
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-0-2-23-lands-an-ordering-fix-75ad97/","markdown":"https://newruntime.com/posts/kapitsa-0-2-23-lands-an-ordering-fix-75ad97.md","json":"https://newruntime.com/posts/kapitsa-0-2-23-lands-an-ordering-fix-75ad97.json"}
---

# Kapitsa 0.2.23 lands an ordering fix aimed at Nemotron-H garbage output

## Retrieval answer

Kapitsa 0.2.23 changes the Flint fused-decode path so merges arrive in the order expected by Nemotron-H.

## Summary
Kapitsa 0.2.23 changes the Flint fused-decode path so merges arrive in the order expected by Nemotron-H. Maintainers call that ordering the likely cause of reported garbage output, while saying M5 Max logits and defaults are unchanged.

For teams running local Apple Silicon inference, this is a focused reliability patch rather than a new capability. It shows how a runtime-ordering defect can be mistaken for a model-quality problem.

## What happened
Kapitsa released version 0.2.23 on October 27. The GitHub release lists three changes: documentation for an 8-bit group-128 requirement, a Flint fused-decode ordering change for Nemotron-H, and version updates across the repository and packaging files. The release was published by the project’s GitHub Actions account and notes 29 commits to `main` since the release.

The substantive runtime change is in Flint, the project’s fused decode implementation. According to the release notes, fused-decode merges now arrive in Nemotron-H’s order. More specifically, thread 0’s fences are placed after the threadgroup barrier. The maintainers describe this ordering as the likely cause of the “garbage” output reported in issue 421. The wording is important: the release attributes the behavior to a likely cause, not to a conclusive root-cause analysis published in the release itself.

The release also says that logits on an M5 Max are bitwise unchanged and that defaults are unchanged. That establishes a limited regression statement for the checked configuration: this patch is not presented as changing default settings or as altering the recorded M5 Max logits. It does not establish equivalent behavior across other Apple hardware, model variants, prompts, quantizations, or runtime configurations. Nor does the release provide benchmark throughput, latency, memory, or quality figures.

A smaller documentation change states that 8-bit group-128 weights need bf16 scales and biases. The release does not expand on which conversion or loading workflow this applies to, nor does it give a migration procedure. Finally, the version bump was propagated through `CMakeLists.txt`, version fallbacks in `src/`, an `install.sh` example, and the Homebrew formula at `packaging/homebrew/kapitsa.rb`. That packaging work makes the 0.2.23 tag a coherent release artifact rather than only a source-level patch.

## What the archive adds
This is a new fact on a known Kapitsa story. The archive covered [Kapitsa v0.2.15](https://newruntime.com/posts/kapitsa-v0-2-15-an-open-source-apple-e8f061/) as an Apple Silicon inference release with stability fixes and faster MoE prefill. Version 0.2.23 narrows the focus to output ordering in the Flint decode path for Nemotron-H.

## Trend signal
This release weakly supports [harness architecture outlives model choice](https://newruntime.com/patterns/harness-architecture-outlives-model-choice/). The model’s expected output can be undermined or restored by runtime-level scheduling and synchronization details, even when a checked set of logits remains bitwise unchanged.

## Why it matters
A model name is not a sufficient description of an inference system’s behavior. In this case, Kapitsa’s release notes point to ordering in a fused decode implementation as the likely source of visibly bad Nemotron-H output. Teams evaluating local model quality should retain the runtime version, quantization details, hardware, and reproducible prompt artifacts alongside model identifiers. Otherwise, a runtime defect can be misclassified as a model limitation.

For current Kapitsa users, the update is also a reminder to treat a release note’s scope precisely. The stated validation is that M5 Max logits are bitwise unchanged and defaults are unchanged; it is not a performance claim or a universal correctness guarantee. Users affected by garbage output have a specific upgrade candidate to test. Users who were not affected have a narrowly scoped maintenance change, not evidence that all inference behavior has changed.

## What happens next
- **Kapitsa will clarify or close the reported garbage-output issue with additional reproduction detail or validation.** (3 months, 55%). Watch for: an issue update, regression test, or follow-up release note.
- **A subsequent release will add more model- or hardware-specific validation around the Flint path.** (6 months, 45%). Watch for: release notes naming additional hardware or Nemotron-H configurations.

## Limits and uncertainty
The public release notes are brief. They do not include the affected Nemotron-H version, a minimal reproduction, before-and-after generated outputs, or a linkable explanation of issue 421 beyond its number. “Likely cause” is explicitly provisional language. The M5 Max statement covers bitwise unchanged logits, but the source does not state the test workload, precision, prompt set, batch shape, or whether the corrected ordering was independently verified across devices. The 8-bit group-128 note gives a requirement but not a full compatibility matrix. This Field Note therefore treats the release as a focused fix claim by the maintainers, not as a broad benchmark or correctness certification.

## Skill to build now
When testing a local inference upgrade, create a small regression bundle: exact model and quantization identifiers, device, runtime version, decoding settings, prompts, raw outputs, and a pass/fail check for expected structure. Run it before and after the update. For systems using fused kernels or model-specific execution paths, include malformed-output checks rather than relying only on aggregate latency or token-per-second measurements. That makes it easier to distinguish a model change from a runtime-ordering failure.

## Reader takeaway
Kapitsa 0.2.23 is a targeted maintenance release: it changes Flint’s fused-decode merge ordering for Nemotron-H to address a likely source of garbage output. Upgrade and test if that symptom matches your setup, but do not read the release as a general performance or quality claim.

## Open question
Will the project publish a reproducible regression test and identify which Nemotron-H configurations and Apple Silicon devices were affected by the ordering issue?

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: [Github](https://newruntime.com/topics/github.md): Explore the github topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)