Thinking Machines Lab says it has trained Inkling 2 Small with the forward pass computed in four-bit numbers, and says no larger model it knows of was trained that way from the first step to the last. The company describes the recipe in a post on its Connectionism blog. Against a full MXFP8 baseline (an 8-bit format), the lab reports that pretraining ran 8% faster by the clock and reinforcement learning (RL) ran 20% faster, measured on its largest jobs.

The trade-off behind all of this is simple. Fewer bits per number means less memory and faster math, but also coarser rounding. Store 3.2 in a format that only holds 3 and 4, and you lose 0.2. One such error is harmless. Millions of them, repeated across layers and optimizer steps, can stop a model from learning well.

Thinking Machines works with three formats: BF16, MXFP8, and NVFP4. NVFP4 takes 3.6 times fewer bits than BF16 yet runs up to 6 times faster on NVIDIA’s GB300 chips, which is where the lab trains. NVFP4 also adds the most quantization error, so the lab does not apply it everywhere.

The split follows where speed matters and where bias hurts. Pretraining chews through trillions of tokens, and each RL rollout generates tens of thousands, so those steps reward speed. Backward passes, the ones that adjust the weights, mix together many tokens in a batch, so random rounding noise washes out. Bias is the danger there: values that consistently round the same direction compound across layers and steps.

So the recipe uses NVFP4 for the forward pass in the expert layers of its mixture-of-experts design, where only some of the network’s parts handle each token. Those layers are expensive to compute and, the lab says, tend to lose less accuracy when quantized. The backward pass runs in MXFP8. Tensors from the forward pass are converted back to higher precision, transposed, then re-quantized with stochastic rounding, which rounds up or down at random in proportion to distance. That keeps the expected value intact and removes systematic bias.

A second fix came from an unexpected place. Once the math got faster, moving data between GPUs became the bottleneck. The lab sends tokens to experts in NVFP4, and compresses the returning results to an 8-bit format after a random Hadamard transform, a rotation that smooths out outliers. That roughly halved the traffic: 16 bits down to 8.25 for each number. Thinking Machines also built its own kernels, codenamed BBBK, to keep the chip’s compute units busy.

The costs are disclosed. Full NVFP4 training hurt model quality more than the lab would accept, and a coarser 16 by 16 scaling approach did too. The last 16.6% of layers stay in MXFP8. In pretraining, the recipe gives up roughly 2% of compute efficiency. Put differently, an MXFP8 run would hit the same loss with about 2% fewer operations. In RL, the lab saw no discernible difference in reward.

These are the lab’s own results on its own model, with no independent replication. Even so, the stake is plain: training time is the scarcest input at frontier labs, and an 8% to 20% cut on the same hardware is worth real money at that scale. Months ago, a run of similar size took several weeks, while Inkling 2 Small was pretrained in 68 hours, though the lab credits many improvements, not numerics alone.

Next on the lab’s list is FP4 for the backward pass and tuning for NVIDIA’s Vera Rubin systems. Any team renting Blackwell-class GPUs now has a published reference for what a 4-bit forward pass costs and what it returns.

Thinking Machines Lab, “End-to-End 4-Bit Training,” Connectionism blog, November 2026.