Inkling 2 Small trained with native 4-bit forward pass on GB300s
The striking part is the duration: 68 hours of pretraining replaces what normally takes weeks, suggesting quantisation-aware training has matured beyond inference-only tricks. Engineers running large-scale training on NVIDIA clusters can study the recipe for memory and throughput gains, and enterprises weighing fine-tuning budgets should watch how it affects cost curves. It also signals that Thinking Machines is competing on infrastructure efficiency rather than model size alone.
Mira Murati announced that her company Thinking Machines has detailed the training of Inkling 2 Small in a new Connectionism blog post. The model was trained end to end with a native 4-bit forward pass on NVIDIA GB300 hardware, spanning pretraining from the first step through the final reinforcement learning rollout. According to the announcement, this and other stack-level improvements compressed pretraining from several weeks down to 68 hours.
The technical approach is notable because 4-bit precision is applied natively across the full training pipeline rather than only at inference or post-training quantisation time. That means the forward pass runs in 4-bit throughout pretraining and RL, which reduces memory traffic and compute cost on the GB300 accelerators.
If the full write-up confirms the numbers, it would place Thinking Machines among the first labs to publish a complete low-bit pretraining recipe at frontier scale, with direct implications for training cost and hardware efficiency for teams building on NVIDIA platforms.
Today on Connectionism: how we trained Inkling 2 Small with a native 4-bit forward pass on @nvidia GB300s, from the first pretraining step through the final RL rollout, one of the many improvements across the stack that took pretraining from weeks to 68 hours.
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Inkling 2 Small, Thinking Machines Lab, NVIDIA GB300.
Source attribution: curated from Mira Murati (TML) (@miramurati).
