Prime Intellect changes its AI training default to curb unstable runs

Prime Intellect released prime-rl 0.10.0 and verifiers 0.4.0 on October 19, putting a new safeguard against unstable AI training into its default settings. The company’s evidence is a demanding software-engineering training test: its new method kept running where a competing approach collapsed. That result comes with an important qualification—the test was deliberately extreme.

The safeguard is INTELLECT Policy Optimization, or IPO, a rule for deciding which generated tokens should contribute to training. It replaces prime-rl’s previous DPPO-based default. IcePop, the method used for the stability comparison, remains available alongside PPO and CISPO for users who explicitly select them.

Filtering disagreement, not just large ratios

In asynchronous reinforcement learning, the system generating responses and the system updating the model can assign different probabilities to the same token—a small unit of text. The model has changed since the response was generated, and the two systems use different computation routines. IPO targets that disagreement rather than assuming their probability estimates match.

IcePop filters tokens when the ratio between the two estimates falls outside a fixed band. Prime illustrates the problem with a rare token whose probability rises from 0.01 to 0.06: that is a sixfold increase, but only a small absolute shift. Conversely, a likely token dropping from 0.90 to 0.20 has changed substantially despite its less dramatic ratio.

IPO instead filters a token when the absolute probability difference exceeds a threshold. It treats increases and decreases alike, regardless of whether the sampled action performed better or worse than expected. The criterion comes from DPPO research; Prime applies it specifically to disagreement between training and response generation.

Prime tested the change on software-engineering reinforcement learning with GLM-4.5-Air, allowing responses to lag the current model by up to 64 training steps. IcePop collapsed around step 1,200; IPO continued at thresholds of 0.2 and 0.3. Prime says IcePop is stable in most runs, so this comparison demonstrates resilience under stress rather than a routine failure of the alternative.

Faster training, with specific comparison conditions

The release also includes a maximum-throughput configuration, distinct from defaults that prioritize fitting models into GPU memory. Prime compared supervised fine-tuning—training on supplied examples—of Qwen3-235B-A22B against torchtitan. Both runs used the same four nodes, each with eight B300 GPUs, back to back, with matching sequence lengths and training settings.

Prime’s same-hardware Qwen3 training comparison

013,732 tokens/s/GPU

prime-rl

Prime reports 4.39 seconds per training step in its Qwen3-235B-A22B test.

023,253 tokens/s/GPU

torchtitan

The comparison run took 5.04 seconds per training step on the same hardware.

Prime attributes most of that gain to changes in mixture-of-experts computation, where selected model components handle each token. Those changes include combining operations and compiling larger blocks together. For GLM-5.3, it also introduced faster attention routines; on real weights and data, the company reports a training step falling from 31.2 to 24.2 seconds.

More models, one evaluation entry point

• New model support includes DeepSeek V4 Flash, Qwen3.8 Flash Next, and Nemotron-H Nano and Super.
• Multimodal reinforcement learning supports Qwen-VL and Kimi K2.5, keeping images inline through response generation, packing and transport rather than sending pixel tensors separately.

The evaluation command moves from verifiers into prime-rl. The unified evaluation stack supports standalone tests and evaluations during training. Interrupted runs can resume from saved traces, generating only missing responses. Shared task selection also means prime-rl and vf-eval choose the same tasks for the same configuration, keeping those two evaluation paths aligned.

There are concrete compatibility changes, too. Verifiers 0.4.0 removes the legacy v0 stack and changes command names to the vf-* prefix. Timed-out responses now count as their own evaluation outcome and stay out of training batches—a distinction between measuring a failed attempt and using it to update the model.