---
type: "post"
id: "a-habr-analysis-argues-that-prismml-s-bonsai-eb9001"
slug: "a-habr-analysis-argues-that-prismml-s-bonsai-eb9001"
title: "A Habr analysis argues that PrismML's Bonsai is not simply a conventional quantization of Qwen weights"
description: "Thesis: Bonsai's reported one-bit model is better understood as a trained low-bit network than as a conventional quantized copy of its source model."
retrieval_nugget: "Thesis: Bonsai's reported one-bit model is better understood as a trained low-bit network than as a conventional quantized copy of its source model."
published_at: "2026-10-18"
updated_at: "2026-10-18"
record_date: "2026-10-18"
date_kind: "published_at"
topics: ["models","local-models"]
source_urls: ["https://habr.com/ru/articles/1095632?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:a-habr-analysis-argues-that-prismml-s-bonsai-eb9001"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:57df259b8e72e92a31eb8eb90cc2b03f842ee939ddfccece9b77c907f983d367","reviewed_at":"2026-10-18T14:04:24.717Z","source_evidence_count":1,"verified_claim_count":3,"site_analysis_schema_version":"newruntime-site-analysis-v1","site_object_kind":"field_note","observed_fact_count":3,"implication_count":1,"watch_condition_count":1,"related_record_count":0}
analysis: {"schema_version":"newruntime-site-analysis-v1","object_kind":"field_note","thesis":"Bonsai's reported one-bit model is better understood as a trained low-bit network than as a conventional quantized copy of its source model.","observed_facts":[{"text":"The article says 27 billion parameters would occupy about 4 GB with one bit per weight plus a 16-bit scale per 128-weight group, versus about 54 GB in FP16, excluding the vocabulary table.","source_urls":["https://habr.com/ru/articles/1095632?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]},{"text":"The author's layerwise solver produced much worse perplexity than Bonsai and had layer errors two to three times higher on unseen text than on calibration text.","source_urls":["https://habr.com/ru/articles/1095632?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]},{"text":"The author reports a 70% match in most likely next token between Bonsai and Qwen in the comparisons, despite low hidden-state cosine similarity.","source_urls":["https://habr.com/ru/articles/1095632?amp%3Butm_medium=rss&amp%3Butm_source=habrahabr"]}],"mechanism":"The author infers from weight and output comparisons that the block weights were trained in a one-bit representation to preserve model outputs, rather than produced by conventional quantization that keeps weights close to the original matrices.","why_now":"Public formats and kernels make one-bit deployment tangible, while the analysis indicates that a storage format alone does not explain a usable model.","implications":["Low-bit deployment may require training methods designed for the representation, not only a more aggressive quantizer."],"evidence_boundary":"This is an independent analysis of weights and outputs; the author did not inspect PrismML's code, so the training method is inferred, not confirmed.","watch_conditions":["Whether PrismML or an independent reproduction publishes training details and out-of-distribution evaluations."],"related_records":[],"new_branch_reason":"No prior New Runtime record is supplied; this investigation starts a distinct analytical record on trained one-bit weights and the evidence boundary around reconstructing Bonsai's method."}
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/a-habr-analysis-argues-that-prismml-s-bonsai-eb9001/","markdown":"https://newruntime.com/posts/a-habr-analysis-argues-that-prismml-s-bonsai-eb9001.md","json":"https://newruntime.com/posts/a-habr-analysis-argues-that-prismml-s-bonsai-eb9001.json"}
---

# A Habr analysis argues that PrismML's Bonsai is not simply a conventional quantization of Qwen weights

## Retrieval answer

Thesis: Bonsai's reported one-bit model is better understood as a trained low-bit network than as a conventional quantized copy of its source model.

Thesis: Bonsai's reported one-bit model is better understood as a trained low-bit network than as a conventional quantized copy of its source model. The article reports that a 27-billion-parameter model would take about 4 GB with one bit per weight and one 16-bit scale per 128-weight group, compared with about 54 GB in FP16, excluding the vocabulary table.

Mechanism: The author infers from weight and output comparisons that the block weights were trained in a one-bit representation to preserve model outputs, rather than produced by conventional quantization that keeps weights close to the original matrices. The author compared Bonsai weights with Qwen3-1.7B and built a layerwise binary solver. Its calibration-layer error improved, but evaluation on unseen text showed layer errors two to three times higher; the assembled model's reported perplexity was 923,707, versus 16.53 for Bonsai. The author also reports that the most likely next token matched between Bonsai and Qwen in 70% of comparisons, despite low hidden-state cosine similarity. These observations support the inference that the model's weights were trained to preserve outputs rather than kept close to the original matrices.

Why now: Public formats and kernels make one-bit deployment tangible, while the analysis indicates that a storage format alone does not explain a usable model. Implication: Low-bit deployment may require training methods designed for the representation, not only a more aggressive quantizer. Evidence boundary: This is an independent analysis of weights and outputs; the author did not inspect PrismML's code, so the training method is inferred, not confirmed. Watch condition: Whether PrismML or an independent reproduction publishes training details and out-of-distribution evaluations.

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: [Models](https://newruntime.com/topics/models.md): Explore the models topic hub.
- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models 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)