Evangelos Georganas and coauthors claim BITCOS losslessly exploits ternary-weight zero density to reach 1.485 bits per weight and improve decode throughput by up to 1.18× on CPUs and 1.27× on tested GPUs, potentially reducing local-inference memory and bandwidth costs.

state: seedheat: lowuncertainty: mediumconvergesscott: lowternary-models quantization local-inference inference-economicsEvangelos GeorganasAlexander HeineckePradeep Dubey

What is this?

The case describes BITCOS as a lossless storage scheme for ternary LLM weights attributed to Evangelos Georganas and coauthors, using a presence bitmap and compacted signs to exploit zero weights. It reports 1.485 bits per weight and decode-throughput gains of up to 1.18× on CPUs and 1.27× on tested GPUs, but the supplied web snippets neither identify BITCOS nor verify its authorship, mechanism, or measurements. The search results provide only a Georganas scholar profile and separate material discussing inference memory-bandwidth bottlenecks; the claimed BITCOS results remain uncorroborated here.

Why it matters to Scott

BITCOS’s claimed joint reduction in weight storage and decode cost aligns with Scott’s Hardware-aware local inference approach, but the hits establish neither ternary-model use on gamepc nor BITCOS compatibility with his Ollama serving path, so it does not yet change a demonstrated deployment decision. The supplied grounding leaves the results uncorroborated; the radar’s B3S base-3 GGUF packing and Falcon3 ternary Triton pages track related developments, not BITCOS itself.

dev:concept.hardware-aware-local-inferencedev:project.gamepcdev:technology.ollamaradar:base3-ternary-gguf-packingradar:falcon3-ternary-triton-speedup

queries asked of Scott's wikis

ternary quantization lossless weight packing zero sparsity
local inference memory bandwidth bottlenecks decode throughput
compressed weights CPU GPU kernel decompression overhead
local model deployment memory footprint inference economics
low-bit model benchmarks hardware-specific speedup claims

Measured heat

now 0 pts/hpeak 0 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 674h

points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion

How the heat travelled

09-13 14:00	⭐ origin echo-reconstructed	BITCOS uses a presence bitmap and compacted sign vector costing 2 − z bits per weight, improves storage over five-trit packing in 26 of 29 t Evangelos Georganas, Alexander Heinecke, and Pradeep Dubey on paper (echo) · attributed from hn.story.49732931	—
09-16 20:59	first on hacker news · published · +79.0h	Breaking the 1.58-bit Barrier for Ternary LLMs matt_d	—
09-16 20:59	amplified on hacker news 👑	hn.story.49732931 matt_d	peak 245 · 41 comments · 100% of case engagement
09-16 21:20	our radar first saw it · +79.3h	discovery anchor: hn.story.49732931	—

pace: p79 vs 1032 stories at the 336h mark (now 674h old) — ahead of openai-gpt-live-1-api (1.0x), behind nsa-ai-testing-billions (1.0x)

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟧 hn	Breaking the 1.58-bit Barrier for Ternary LLMs Retrieved article excerpt Open article · Retrieved 2026-09-16T21:22:30.620782+00:00 # Computer Science > Artificial Intelligence arXiv:2609.16338 (cs) [Submitted on 14 Sep 2026] # Title:Breaking the 1.58-bit Barrier for Ternary LLMs Authors:[Evangelos Georganas](https://arxiv.org/search/cs?searchtype=author&query=Georganas,+E), [Alexander Heinecke](https://arxiv.org/search/cs?searchtype=author&query=Heinecke,+A), [Pradeep Dubey](https://arxiv.org/search/cs?searchtype=author&query=Dubey,+P) View a PDF of the paper titled Breaking the 1.58-bit Barrier for Ternary LLMs, by Evangelos Georganas and 2 other authors [View PDF](https://arxiv.org/pdf/2609.16338) [HTML (experimental)](https://arxiv.org/html/2609.16338v1) > Abstract:Ternary Large Language Models (LLM) store every weight as one of three symbols $\{-1,0,+1\}$, so the cost of a ternary model is conventionally referenced to the information-theoretic $\log\_2 3 \approx 1.585$ bits per weight. The prevailing deployment format packs five ternary weights into one byte (five-trit packing), and due to the power-of-two group sizes used in practice this rounds up to $1.625$ bits per weight. This effective storage bit-width treats the three symbols $\{-1,0,+1\}$ as equiprobable. We measure the actual symbol distribution of 29 ternary LLM models and find that zeros account for up to $51.5\%$ of all weights. Motivated by this finding, we introduce BITCOS, a simple distribution-adaptive layout comprised of a dense presence bitmap plus a compacted sign vector, and costs $2 - z$ bits per weight element given a zero density $z$ in the model's weights. BITCOS stores weights more compactly than the five-trit packing in 26 of the 29 tested models, and reaches $1.485$ bits per weight on the sparsest of them. BITCOS is amenable to efficient unpacking on modern processors and GPUs, and we present optimized unpacking sequences for AVX-512, AVX2 and Intel Xe2 GPUs. Measured against production state-of-the-art ternary matrix-vector multiplication kernels, at the zero densities real-world ternary models exhibit, the realized gain with our proposed layout is up to $1.28\times$. Finally, we illustrate end-to-end LLM inference results on 5 different platforms (client and server CPUs, integrated and discrete Xe2 GPUs) where decode throughput improves by up to $1.18\times$ on CPUs and $1.27\times$ on GPUs. \| \| \| \| --- \| --- \| \| Subjects: \| Artificial Intelligence (cs.AI); Machine Learning (cs.LG) \| \| Cite as: \| [arXiv:2609.16338](https://arxiv.org/abs/2609.16338) [cs.AI] \| \| \| (or [arXiv:2609.16338v1](https://arxiv.org/abs/2609.16338v1) [cs.AI] for this version) \| \| \| <https://doi.org/10.48550/arXiv.2609.16338> Focus to learn more arXiv-issued DOI via DataCite (pending registration) \| ## Submission history From: Evangelos Georganas [[view email](https://arxiv.org/show-email/6908fb0d/2609.16338)] [v1] Mon, 14 Sep 2026 20:54:24 UTC (144 KB) Full-text links: ## Access Paper: View a PDF of the paper titled Breaking the 1.58-bit Barrier for Ternary LLMs, by Evangelos Georganas and 2 other authors - [View PDF](https://arxiv.org/pdf/2609.16338) - [HTML (experimental)](https://arxiv.org/html/2609.16338v1) - [TeX Source](https://arxiv.org/src/2609.16338) [license icon](http://creativecommons.org/licenses/by/4.0/ "Rights to this article") ### Current browse context: cs.AI [< prev](https://arxiv.org/prevnext?id=2609.16338&function=prev&context=cs.AI "previous in cs.AI (accesskey p)") \| [next >](https://arxiv.org/prevnext?id=2609.16338&function=next&context=cs.AI "next in cs.AI (accesskey n)") [new](https://arxiv.org/list/cs.AI/new) \| [recent](https://arxiv.org/list/cs.AI/recent) \| [2026-09](https://arxiv.org/list/cs.AI/2026-09) Change to browse by: [cs](https://arxiv.org/abs/2609.16338?context=cs) [cs.LG](https://arxiv.org/abs/2609.16338?context=cs.LG) ### References & Citations - [NASA ADS](https://ui.adsabs.harvard.edu/abs/arXiv:2609.16338) - [Google Scholar](https://scholar.google.com/scholar_lookup?arxiv_id=2609.16338) - [Semantic Scholar](https://api.semanticscholar.org/arXiv:2609.16338) export BibTeX citation Loading... ## BibTeX formatted citation × loading... Data provided by: ### Bookmark [BibSonomy](http://www.bibsonomy.org/BibtexHandler?requTask=upload&url=https://arxiv.org/abs/2609.16338&description=Breaking the 1.58-bit Barrier for Ternary LLMs "Bookmark on BibSonomy") [Reddit](https://reddit.com/submit?url=https://arxiv.org/abs/2609.16338&title=Breaking the 1.58-bit Barrier for Ternary LLMs "Bookmark on Reddit") Bibliographic Tools # Bibliographic and Citation Tools Bibliographic Explorer Toggle Bibliographic Explorer ([What is the Explorer?](https://info.arxiv.org/labs/showcase.html#arxiv-bibliographic-explorer)) Connected Papers Toggle Connected Papers ([What is Connected Papers?](https://www.connectedpapers.com/about)) Litmaps Toggle Litmaps ([What is Litmaps?](https://www.litmaps.co/)) scite.ai Toggle scite Smart Citations ([What are Smart Citations?](https://www.scite.ai/)) Code, Data, Media # Code, Data and Media Associated with this Article alphaXiv Toggle alphaXiv ([What is alphaXiv?](https://alphaxiv.org/)) Links to Code Toggle CatalyzeX Code Finder for Papers ([What is CatalyzeX?](https://www.catalyzex.com)) DagsHub Toggle DagsHub ([What is DagsHub?](https://dagshub.com/)) GotitPub Toggle Gotit.pub ([What is GotitPub?](http://gotit.pub/faq)) Huggingface Toggle Hugging Face ([What is Huggingface?](https://huggingface.co/huggingface)) ScienceCast Toggle ScienceCast ([What is ScienceCast?](https://sciencecast.org/welcome)) Demos # Demos Replicate Toggle Replicate ([What is Replicate?](https://replicate.com/docs/arxiv/about)) Spaces Toggle Hugging Face Spaces ([What is Spaces?](https://huggingface.co/docs/hub/spaces)) Spaces Toggle TXYZ.AI ([What is TXYZ.AI?](https://txyz.ai)) Related Papers # Recommenders and Search Tools Link to Influence Flower Influence Flower ([What are Influence Flowers?](https://influencemap.cmlab.dev/)) Core recommender toggle CORE Recommender ([What is CORE?](https://core.ac.uk/services/recommender)) - Author - Venue - Institution - Topic About arXivLabs # arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? [Learn more about arXivLabs](https://info.arxiv.org/labs/index.html). [Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2609.16338) \| Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))	matt_d	245	41
🟧 echo.paper ⭐	BITCOS uses a presence bitmap and compacted sign vector costing 2 − z bits per weight, improves storage over five-trit packing in 26 of 29 t	Evangelos Georganas, Alexander Heinecke, and Pradeep Dubey	—	—

source

object

author

score

comments

🟧 hn

Breaking the 1.58-bit Barrier for Ternary LLMs

Retrieved article excerpt

Open article · Retrieved 2026-09-16T21:22:30.620782+00:00

# Computer Science > Artificial Intelligence

**arXiv:2609.16338** (cs)

[Submitted on 14 Sep 2026]

# Title:Breaking the 1.58-bit Barrier for Ternary LLMs

Authors:[Evangelos Georganas](https://arxiv.org/search/cs?searchtype=author&query=Georganas,+E), [Alexander Heinecke](https://arxiv.org/search/cs?searchtype=author&query=Heinecke,+A), [Pradeep Dubey](https://arxiv.org/search/cs?searchtype=author&query=Dubey,+P)

View a PDF of the paper titled Breaking the 1.58-bit Barrier for Ternary LLMs, by Evangelos Georganas and 2 other authors

[View PDF](https://arxiv.org/pdf/2609.16338)
[HTML (experimental)](https://arxiv.org/html/2609.16338v1)
> Abstract:Ternary Large Language Models (LLM) store every weight as one of three symbols $\{-1,0,+1\}$, so the cost of a ternary model is conventionally referenced to the information-theoretic $\log\_2 3 \approx 1.585$ bits per weight. The prevailing deployment format packs five ternary weights into one byte (five-trit packing), and due to the power-of-two group sizes used in practice this rounds up to $1.625$ bits per weight. This effective storage bit-width treats the three symbols $\{-1,0,+1\}$ as equiprobable. We measure the actual symbol distribution of 29 ternary LLM models and find that zeros account for up to $51.5\%$ of all weights. Motivated by this finding, we introduce BITCOS, a simple distribution-adaptive layout comprised of a dense presence bitmap plus a compacted sign vector, and costs $2 - z$ bits per weight element given a zero density $z$ in the model's weights. BITCOS stores weights more compactly than the five-trit packing in 26 of the 29 tested models, and reaches $1.485$ bits per weight on the sparsest of them. BITCOS is amenable to efficient unpacking on modern processors and GPUs, and we present optimized unpacking sequences for AVX-512, AVX2 and Intel Xe2 GPUs. Measured against production state-of-the-art ternary matrix-vector multiplication kernels, at the zero densities real-world ternary models exhibit, the realized gain with our proposed layout is up to $1.28\times$. Finally, we illustrate end-to-end LLM inference results on 5 different platforms (client and server CPUs, integrated and discrete Xe2 GPUs) where decode throughput improves by up to $1.18\times$ on CPUs and $1.27\times$ on GPUs.

|  |  |
| --- | --- |
| Subjects: | Artificial Intelligence (cs.AI); Machine Learning (cs.LG) |
| Cite as: | [arXiv:2609.16338](https://arxiv.org/abs/2609.16338) [cs.AI] |
|  | (or  [arXiv:2609.16338v1](https://arxiv.org/abs/2609.16338v1) [cs.AI] for this version) |
|  | <https://doi.org/10.48550/arXiv.2609.16338> Focus to learn more  arXiv-issued DOI via DataCite (pending registration) |

## Submission history

From: Evangelos Georganas [[view email](https://arxiv.org/show-email/6908fb0d/2609.16338)]   
 **[v1]**
Mon, 14 Sep 2026 20:54:24 UTC (144 KB)

Full-text links:

## Access Paper:

View a PDF of the paper titled Breaking the 1.58-bit Barrier for Ternary LLMs, by Evangelos Georganas and 2 other authors

- [View PDF](https://arxiv.org/pdf/2609.16338)
- [HTML (experimental)](https://arxiv.org/html/2609.16338v1)
- [TeX Source](https://arxiv.org/src/2609.16338)

[license icon](http://creativecommons.org/licenses/by/4.0/ "Rights to this article")

### Current browse context:

cs.AI

[< prev](https://arxiv.org/prevnext?id=2609.16338&function=prev&context=cs.AI "previous in cs.AI (accesskey p)")
  |   
[next >](https://arxiv.org/prevnext?id=2609.16338&function=next&context=cs.AI "next in cs.AI (accesskey n)")

[new](https://arxiv.org/list/cs.AI/new)
 | 
[recent](https://arxiv.org/list/cs.AI/recent)
 | [2026-09](https://arxiv.org/list/cs.AI/2026-09)

Change to browse by:

[cs](https://arxiv.org/abs/2609.16338?context=cs)  
[cs.LG](https://arxiv.org/abs/2609.16338?context=cs.LG)

### References & Citations

- [NASA ADS](https://ui.adsabs.harvard.edu/abs/arXiv:2609.16338)
- [Google Scholar](https://scholar.google.com/scholar_lookup?arxiv_id=2609.16338)
- [Semantic Scholar](https://api.semanticscholar.org/arXiv:2609.16338)

export BibTeX citation
Loading...

## BibTeX formatted citation

×

loading...

Data provided by:

### Bookmark

[BibSonomy](http://www.bibsonomy.org/BibtexHandler?requTask=upload&url=https://arxiv.org/abs/2609.16338&description=Breaking the 1.58-bit Barrier for Ternary LLMs "Bookmark on BibSonomy")
[Reddit](https://reddit.com/submit?url=https://arxiv.org/abs/2609.16338&title=Breaking the 1.58-bit Barrier for Ternary LLMs "Bookmark on Reddit")



Bibliographic Tools

# Bibliographic and Citation Tools

Bibliographic Explorer Toggle

Bibliographic Explorer *([What is the Explorer?](https://info.arxiv.org/labs/showcase.html#arxiv-bibliographic-explorer))*

Connected Papers Toggle

Connected Papers *([What is Connected Papers?](https://www.connectedpapers.com/about))*

Litmaps Toggle

Litmaps *([What is Litmaps?](https://www.litmaps.co/))*

scite.ai Toggle

scite Smart Citations *([What are Smart Citations?](https://www.scite.ai/))*

Code, Data, Media

# Code, Data and Media Associated with this Article

alphaXiv Toggle

alphaXiv *([What is alphaXiv?](https://alphaxiv.org/))*

Links to Code Toggle

CatalyzeX Code Finder for Papers *([What is CatalyzeX?](https://www.catalyzex.com))*

DagsHub Toggle

DagsHub *([What is DagsHub?](https://dagshub.com/))*

GotitPub Toggle

Gotit.pub *([What is GotitPub?](http://gotit.pub/faq))*

Huggingface Toggle

Hugging Face *([What is Huggingface?](https://huggingface.co/huggingface))*

ScienceCast Toggle

ScienceCast *([What is ScienceCast?](https://sciencecast.org/welcome))*

Demos

# Demos

Replicate Toggle

Replicate *([What is Replicate?](https://replicate.com/docs/arxiv/about))*

Spaces Toggle

Hugging Face Spaces *([What is Spaces?](https://huggingface.co/docs/hub/spaces))*

Spaces Toggle

TXYZ.AI *([What is TXYZ.AI?](https://txyz.ai))*

Related Papers

# Recommenders and Search Tools

Link to Influence Flower

Influence Flower *([What are Influence Flowers?](https://influencemap.cmlab.dev/))*

Core recommender toggle

CORE Recommender *([What is CORE?](https://core.ac.uk/services/recommender))*

- Author
- Venue
- Institution
- Topic


About arXivLabs

# arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? [**Learn more about arXivLabs**](https://info.arxiv.org/labs/index.html).

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2609.16338) |
Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))

matt_d

245

🟧 echo.paper ⭐

BITCOS uses a presence bitmap and compacted sign vector costing 2 − z bits per weight, improves storage over five-trit packing in 26 of 29 t

Evangelos Georganas, Alexander Heinecke, and Pradeep Dubey

—

Interpretation history

2026-09-17T11:27:45Z
The file-format-only concern is an unsupported question, not a credible contradiction: the supplied abstract explicitly reports optimized unpacking, matrix-vector kernels, and end-to-end decode gains. Discussion adds no replication or deployment evidence, so the case remains a bounded author-reported optimization rather than an actionable local-inference change.
2026-09-16T21:45:50Z
grounded: converges/low — BITCOS’s claimed joint reduction in weight storage and decode cost aligns with Scott’s Hardware-aware local inference approach, but the hits establish neither t
2026-09-16T21:40:20Z
case created — The paper supplies a bounded storage-layout and runtime claim distinct from existing ternary packing episodes, not a new lossy quantization method.

Decision trace

10-08 00:51review_dormantscheduled targets exhausted or 28 quiet days
10-08 00:51drop_targetsquiet through full ladder or over cap 8
09-18 04:35review_screenThe changes add speculative packing suggestions and remove/reorder discussion, without introducing a new fact, implementation result, contradiction, release, or consequential evidence.
09-17 21:27repriceThe file-format-only concern is an unsupported question, not a credible contradiction: the supplied abstract explicitly reports optimized unpacking, matrix-vector kernels, and end-to-end decode gains.
09-17 21:27review_screenA new comment raises a potentially consequential but unverified concern that the compression may apply only to the file format and require expansion in memory, which could qualify the claimed inferenc
09-17 16:20sensor_dirtycomment_update
09-17 14:29review_screenThe new comments reiterate the known entropy/zero-density packing result and offer opinions or speculation, without new implementation evidence, contradiction, release, or access change.
09-17 09:20sensor_dirtycomment_update
09-17 07:45groundBITCOS’s claimed joint reduction in weight storage and decode cost aligns with Scott’s Hardware-aware local inference approach, but the hits establish neither ternary-model use on gamepc nor BITCOS co
09-17 07:40createThe paper supplies a bounded storage-layout and runtime claim distinct from existing ternary packing episodes, not a new lossy quantization method.