Thinking Machines Lab published its end-to-end 4-bit training recipe on its official blog on November 13: running NVFP4 on NVIDIA GB300 hardware, the full pipeline from pretraining to reinforcement learning runs 8% faster in pretraining and 20% faster in RL, at a cost of about 2% in compute efficiency. The key word is "end to end": instead of training in high precision and quantizing the weights afterward, the model is designed for 4-bit from day one, and the 594B-parameter Inkling 2 Small is the largest model known to have been trained this way all the way through.
How NVFP4 runs on the GB300
The recipe mixes three precisions. In the forward pass, the MoE expert layers run in NVFP4, averaging 4.5 bits per value and reaching 15 PFLOPS on the GB300 — 3x MXFP8 and 6x BF16. The backward pass uses MXFP8 with stochastic rounding to keep gradient bias down, and the final 16.6% of layers stay in MXFP8 rather than dropping to 4-bit. Communication between MoE layers is compressed too, squeezed into the S2F6 format via random Hadamard transforms. Thinking Machines Lab says the recipe needs no parameter changes across its largest jobs, from supervised fine-tuning through reinforcement learning.
What the speedup costs
The price is roughly 2% of compute efficiency: the model needs a little more compute to reach the same loss, with no noticeable effect on reward. But all of these numbers come from the lab itself: no code or configuration has been published, and no independent reproduction exists yet. The audience is also narrow — the recipe is tied to Blackwell-class chips like the GB300, so only labs with large clusters can use it; ordinary developers and consumer GPUs are outside its reach.
Two things matter next: whether Thinking Machines Lab publishes the training code and configuration, and whether another lab reproduces similar speedups on the GB300. If reproduction holds up, 4-bit training moves from a single lab's craft to an industry default, and frontier training bills get recalculated; if it fails, this remains a well-written blog post.