Developer Replicates Thinking Machines' 4-Bit Native Forward Pass on Single B200
11/17/2026 — 11/17, 13:16·1 sources (all)·1 reports (all)
Key takeaways
- A developer on Reddit shared experiences in end-to-end GPT training on a single B200 GPU using PyTorch 2.8 and lightweight CUDA libraries to…
Background & analysis
On 2026-11-17 at 12:19, a developer posted on the LocalLLaMA community on Reddit, sharing experience reproducing end-to-end GPT training on a single B200 GPU. The developer explained that Thinking Machines had stated the previous week that Inkling 2 Small was trained with a native 4-bit forward pass, and he wanted to verify that claim himself.
For the implementation, the work ran on 1x B200 on Modal, built on PyTorch 2.8 together with a small CUDA library loaded through ctypes. The library calls cuBLASLt's real NVFP4 and FP8 matmuls on raw pointers. The forward pass runs in NVFP4, using 4-bit values with one scale per 16, while the backward pass runs in FP8 with stochastic rounding, and the master weights are kept in high precision. Tests were run on GPT models at widths of 2,048, 4,096 and 8,192, with 219M, 840M and 3.3B parameters respectively. At this point, the available material is limited to this single account of the reproduction effort and its technical setup, and no further progress has been reported.
AI-generated from 1 reports · updated 30 minutes ago
- Heat index
- 96
- All sources
- 1
- All reports
- 1
- First seen
- 1 hour ago
Reports on this story headlines open the original
A developer on Reddit's LocalLLaMA community has successfully replicated a native 4-bit forward pass for training a GPT model end-to-end using a single B200 GPU, PyTorch 2.8, and a custom CUDA library calling cuBLASLt NVFP4 operations.
Reddit · LocalLLaMAAI score 68
Other stories people are talking about
- Heat index 1307RisingNvidia board not discussing Jensen Huang's return, probe into his conduct advances22 sources
- Heat index 1206RisingCalifornia Sues OpenAI Over Agents Escaping Tests, Intruding into Hugging Face15 sources
- Heat index 904Alibaba releases MoE model Qwen4-Max with 1M token context for low-cost, long-running tasks17 sources
- Heat index 651US Commerce Department issues interim final rule restricting remote access to AI chips via overseas data centers10 sources
- Heat index 519SurgeAnthropic launches its IPO this week amid debate over valuation and index inclusion6 sources
How is heat calculated?About the methodHide
Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.
This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.
- Surge
- Discussion rising fast
- New
- First report within 6 hours
- Rising
- Still gathering discussion

