End‑to‑End 4‑Bit Training
Comments on “End‑to‑End 4‑Bit Training”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Thinking Machines Lab describes how it trained Inkling 2 Small with NVFP4 in the mixture-of-experts forward pass and MXFP8 in the backward pass, using re-quantization with stochastic rounding. It also quantizes expert-parallel communication, including combine, with a Hadamard transform and S2F6. The lab reports an 8% pretraining and 20% RL wall-clock speedup over full MXFP8, at about 2% lost compute efficiency in pretraining and no discernible reward change in RL. These figures are self-reported, from the lab's largest jobs, and have no independent source.
Why it ranked
No. 5 of 123 with a score of 0.75. The panel rated it highest on technical depth (0.97) and new information (0.92), lowest on audience breadth (0.64).
- AI judgeJudge 2claude-haiku-5-5
- Deep, specific engineering writeup on end-to-end NVFP4 LLM training with concrete numerics, kernels and measured speedups; substantive and new, though aimed at a specialized audience.
- AI judgeJudge 3gpt-5.6-luna
- Detailed engineering write-up of end-to-end NVFP4 training, including stochastic re-quantization, communication compression, specialized kernels, measured speedups, accuracy tradeoffs, and reproducible numerical implementation details.
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
0.75total under rubric v2.3
- Significance0.75
- Reader interest0.85
- Audience breadth0.64
- Practical relevance0.76
- Substantiated0.87
- Technical depth0.97
- New information0.92
- Counts against: Promotional0.08
- Counts against: Overclaim0.23
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? | 0.67 | 0.72 | 0.86 | 0.75 | 0.19 | +0.24 | +0.180 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.72 | 0.90 | 0.94 | 0.85 | 0.22 | +0.26 | +0.222 |
| Audience breadthHow many people who build technology does it concern? | 0.52 | 0.70 | 0.70 | 0.64 | 0.18 | +0.13 | +0.083 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.64 | 0.75 | 0.88 | 0.76 | 0.24 | +0.22 | +0.166 |
| 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. | 0.84 | 0.85 | 0.93 | 0.87 | 0.09 | +0.05 | +0.044 |
| Technical depthHow technically substantive is it? | 0.99 | 0.95 | 0.98 | 0.97 | 0.04 | +0.05 | +0.049 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.89 | 0.90 | 0.98 | 0.92 | 0.09 | +0.05 | +0.046 |
| 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. | 0.10 | 0.05 | 0.08 | 0.08 | 0.05 | −0.19 | −0.015 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.38 | 0.20 | 0.12 | 0.23 | 0.26 | −0.11 | −0.026 |
| Sum of contributions | +0.750 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.750 | ||||||
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/2026-11-14/5/"><img src="https://rubric.hrih.si/badge/2026-11-14/5.svg" alt="Ranked #5 on Rubric, 14 Nov 2026" width="238" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenFirst, only or fastest
Inkling 2 Small is the largest model the authors know of that was trained end-to-end with a native 4-bit forward pass.
The article gives no link or citation for it.
The article says “Inkling 2 Small is the largest model we are aware of that was trained end-to-end with a native 4-bit forward pass”
The source says No source to compare with.
No source givenNumber
In the authors' largest jobs, the NVFP4/MXFP8 recipe gives an 8% wall-clock speedup in pretraining and a 20% speedup in RL over full MXFP8 training.
The article gives no link or citation for it.
The article says “this recipe provides us with an 8% wall-clock speedup in pretraining and a 20% speedup in RL over full MXFP8 training”
The source says No source to compare with.
No source givenNumber
NVFP4 is stored with 3.6x fewer bits than BF16 but runs up to 6x faster on GB300s.
The article gives no link or citation for it.
The article says “NVFP4 is stored with 3.6x fewer bits than BF16, but it runs up to 6x faster on GB300s”
The source says No source to compare with.
No source givenGeneralization
Full MXFP8 training of the MoE layers barely affected the model's compute efficiency, whereas initial full NVFP4 training caused very large hits to it.
The article gives no link or citation for it.
The article says “we implemented full MXFP8 training of our MoE layers and found it barely affected our model’s compute efficiency (CE”
The source says No source to compare with.
No source givenCausal conclusion
Quantizing combine-stage activations with a 32-dimensional random Hadamard transform and block-scaled S2F6 resulted in no measurable accuracy degradation, while MXFP8 impaired performance too much.
The article gives no link or citation for it.
The article says “performing a 32-dimensional random Hadamard transform (RHT) and then quantizing to block-scaled S2F6, another 8-bit format, resulted in no measurable accuracy degradation during training”
The source says No source to compare with.
No source givenNumber
The S2F6 combine quantization reduces the data transmitted during combine from 16 bits per value to 8.25 bits per value.
The article gives no link or citation for it.
The article says “This reduces the volume of data transmitted during combine from 16 bits per value to 8.25 bits per value”
The source says No source to compare with.
No source givenNumber
In pretraining Inkling 2 Small, the recipe costs about 2% in compute efficiency relative to MXFP8, and in RL it gave a substantial speedup with no discernible difference in reward.
The article gives no link or citation for it.
The article says “In pretraining Inkling 2 Small it costs about 2% in compute efficiency”
The source says No source to compare with.
No source givenNumber
Inkling 2 Small was pretrained in 68 hours on GB300 NVL72 racks, whereas a similarly sized model took several weeks months earlier.
The article gives no link or citation for it.
The article says “Inkling 2 Small was pretrained in just 68 hours on GB300 NVL72 racks”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 24 pts, front page
Recorded for context only, checked 15 Nov 2026, 20:44 UTC. It played no part in the ranking.