AI News

Trending story

Thinking Machines Lab introduces the Inkling 2 Small end-to-end 4-bit training recipe

3 reports1 discussions4 story updatesUpdated Nov 14 07:32

Understand this story

AI overview

On November 13, 2026, Thinking Machines Lab published a technical blog post introducing the end-to-end native 4-bit training recipe for the Inkling 2 Small model. The approach uses NVFP4 precision in forward computation and MXFP8 in backpropagation, combined with requantization and stochastic rounding, while also quantizing GPU communication data to alleviate bottlenecks. According to the official blog, on the largest-scale tasks this recipe speeds up pretraining by 8% and reinforcement learning by 20%, at the cost of roughly a 2% loss in pretraining compute efficiency; the team says this is the largest end-to-end 4-bit forward training model known to date. Reports subsequently stated that Thinking Machines Lab has open-sourced the 594B-parameter Inkling 2 Small, available for download on HuggingFace under the Apache 2.0 license, requiring at least 500GB of GPU memory (NVFP4) or 800GB (MXFP8) to run.

AI-generated from public reports

Latest development · Report

Thinking Machines Lab Open-Sources the 594B-Parameter Model Inkling 2 Small

Developments

1 confirmed developments

  1. Thinking Machines Lab introduces the end-to-end 4-bit training recipe for Inkling 2 Small

    Source:English text is being prepared.

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 14

  1. Rohan Paul

    Thinking Machines Lab Open-Sources the 594B-Parameter Model Inkling 2 Small

    Thinking Machines Lab has open-sourced Inkling 2 Small, a 594B-parameter model, which is now available for download on Hugging Face under the Apache 2.0 license. The model uses native 4-bit forward computation throughout, from the first step of pretraining to the final RL rollout, shortening pretraining time by 8% and RL time by 20%, with an efficiency loss of about 2%. Running it requires at least 500GB of GPU memory (NVFP4) or 800GB (MXFP8); MXFP8 is only 1–2 percentage points higher than NVFP4 on evaluations. Because the same numerical format is maintained throughout training, the released 4-bit checkpoint requires no separate calibration.

  2. TeortaxesRelated discussion

    Teortaxes joined this discussion

    This post contributes to heat; it is not a formal report or confirmed fact.

  3. Thinking Machines Lab: Official BlogFirst-partySelected

    Thinking Machines Lab details end-to-end 4-bit training: Inkling 2…

    Thinking Machines Lab has published a technical blog post introducing that its Inkling 2 Small model uses end-to-end native 4-bit forward training, achieving an 8% pretraining speedup and a 20% RL speedup on its largest-scale tasks. The approach uses NVFP4 precision in the forward pass and MXFP8 in the backward pass, combined with requantization and stochastic rounding, while also quantizing GPU communication data to alleviate bottlenecks. The team says this is the largest end-to-end 4-bit forward-training model known to date, at the cost of roughly a 2% loss in pretraining compute efficiency.

  4. Mira Murati

    Thinking Machines Details Inkling 2 Small's Native 4-bit Training

    Thinking Machines detailed in Connectionism how it trained Inkling 2 Small using native 4-bit forward computation, covering the entire process from the first step of pretraining to the final RL rollout. The training was conducted on NVIDIA GB300 and is one of several improvements across the full technology stack, reducing pretraining time from several weeks to 68 hours. Mira Murati reposted this content.

Heat trend for this story

Current
5
Period peak
48 Nov 14 08:00
24-hour change
-80%

Move the pointer or select the chart to inspect hourly heat. Use the left and right arrow keys with a keyboard.

Related stories

  1. Related storiesNvidia releases Nemotron 3.5 Super VL and Bolt models
  2. Related storiesMeta releases open-weight model Muse Spark Open
  3. Related storiesSGLang natively supports and powers the Beam production API
  4. Related storiesEnglish translation is being prepared.