Lossless compression of LLMs to 12-bits
Comments on “Lossless compression of LLMs to 12-bits”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
NF12, an open-source library from dphnAI (Apache-2.0), stores BF16 LLM weights in a lossless 12-bit format, following a 2026 Together AI paper, with CUDA kernels for Ampere through Blackwell. Qwen3-8B shrinks from 15.27 to 11.49 GiB, and Qwen3-30B-A3B from 56.88 to 42.80 GiB. The README's throughput results are mixed: up to 1.34× BF16 on Jetson Thor, but 0.81× on B200 at concurrency 32. These figures are self-reported and could not be independently checked.
Why it ranked
No. 12 of 46 with a score of 0.62. The panel rated it highest on new information (0.91) and technical depth (0.81), lowest on audience breadth (0.51). Placed No. 12, where the score alone would also have put it: the fourth AI & machine learning story of the day (−0.020). The placement terms move stories; they are not part of the score.
- AI judgeJudge 2claude-haiku-5-5
- Library release for lossless 12-bit BF16 weight compression with GPU kernels and benchmarks; substantive, practical for LLM inference builders, backed by a paper and measurements, but niche in audience.
- AI judgeJudge 3gpt-5.6-luna
- A concrete GPU library delivers lossless 12-bit LLM weight compression with APIs, compatibility constraints, memory savings, and comparative benchmarks. It is technically substantive and actionable, though broad significance depends on adoption and benchmark reproducibility.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.62total under rubric v2.3
- Significance0.57
- Reader interest0.71
- Audience breadth0.51
- Practical relevance0.75
- Substantiated0.76
- Technical depth0.81
- New information0.91
- Counts against: Promotional0.14
- Counts against: Overclaim0.28
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.55 | 0.45 | 0.72 | 0.57 | 0.27 | +0.24 | +0.138 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.64 | 0.60 | 0.88 | 0.71 | 0.28 | +0.26 | +0.184 |
| Audience breadthHow many people who build technology does it concern? | 0.45 | 0.50 | 0.58 | 0.51 | 0.13 | +0.13 | +0.066 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.78 | 0.65 | 0.82 | 0.75 | 0.17 | +0.22 | +0.165 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.91 | 0.60 | 0.78 | 0.76 | 0.31 | +0.05 | +0.038 |
| Technical depthHow technically substantive is it? | 0.86 | 0.70 | 0.86 | 0.81 | 0.16 | +0.05 | +0.040 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.86 | 0.90 | 0.96 | 0.91 | 0.10 | +0.05 | +0.045 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.09 | 0.20 | 0.12 | 0.14 | 0.11 | −0.19 | −0.026 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.47 | 0.30 | 0.08 | 0.28 | 0.39 | −0.11 | −0.031 |
| Sum of contributions | +0.620 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.620 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-11-09/12/"><img src="https://rubric.hrih.si/badge/2026-11-09/12.svg" alt="Ranked #12 on Rubric, 9 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 4 no source given · 4 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenVersion
NF12 is a lossless, fixed-length 12-bit encoding of BF16 weight matrices, based on the Together AI paper "NyoomFloat12" (2026).
The article gives no link or citation for it.
The article says “NF12 is a lossless, fixed-length 12-bit encoding of BF16 weight matrices, based on "NyoomFloat12: Accelerating LLM Inference via Lossless 12-bit Weight Compression”
The source says No source to compare with.
Couldn’t checkVersion
The library's bit and memory layout (tiled) are not compatible with the layout described in the paper.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “This library's bit and memory layout (tiled) are not compatible with the one described in the paper”
The source says No passage could be compared.
Source cited by the article: github.com/dphnAI/nf12/blob/master/docs/SPEC.md
No source givenVersion
Linux x86-64 and ARM64 wheels include kernels for SM80, 86, 89, 90, 100, 110, and 120.
The article gives no link or citation for it.
The article says “Linux x86-64 and ARM64 wheels include kernels for SM80, 86, 89, 90, 100, 110, and 120”
The source says No source to compare with.
Couldn’t checkBenchmark result
On Jetson Thor (SM110), NF12 reached 1.32× / 1.34× / 1.28× the BF16 output tokens per second on Qwen3-8B at concurrency 1 / 8 / 32.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “Jetson Thor (SM110) 1.32× / 1.34× / 1.28×”
The source says No passage could be compared.
Source cited by the article: github.com/dphnAI/nf12/blob/master/bench
Couldn’t checkBenchmark result
On B200 (SM100), NF12 was slower than BF16 at higher concurrency on Qwen3-8B (0.95× at 8, 0.81× at 32).
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “B200 (SM100) 1.07× / 0.95× / 0.81×”
The source says No passage could be compared.
Source cited by the article: github.com/dphnAI/nf12/blob/master/bench
No source givenNumber
NF12 reduces Qwen3-8B weight size from 15.27 GiB (BF16) to 11.49 GiB.
The article gives no link or citation for it.
The article says “Qwen3-8B 15.27 11.49”
The source says No source to compare with.
No source givenNumber
Qwen3-30B-A3B shrinks from 56.88 GiB in BF16 to 42.80 GiB in NF12.
The article gives no link or citation for it.
The article says “Qwen3-30B-A3B 56.88 42.80”
The source says No source to compare with.
Couldn’t checkBenchmark result
NF12 fused MoE layer latency on B200 is 18.9 µs for Qwen3-30B-A3B, a 1.48× speedup over the previous composed NF12 path.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “B200 18.9 (1.48×) 22.1 (1.49×”
The source says No passage could be compared.
Source cited by the article: github.com/dphnAI/nf12/blob/master/bench
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 9 Nov 2026, 05:46 UTC. It played no part in the ranking.