Morning Edition · Saturday, July 18, 2026Published at 1:46 AM EDT · New York

Li Auto's 35B Mixture-of-Experts Model Claims 100B-Class Performance From Post-Training Alone

Mach-Mind-4-Flash activates 3 billion parameters per token and reaches its results without scaling pretraining compute, using multi-track reinforcement learning (RL), multi-teacher distillation and a token-efficiency stage.

Save

Li Auto's 35B Mixture-of-Experts Model Claims 100B-Class Performance From Post-Training Alone

The AI division of Chinese electric-vehicle maker Li Auto (Lixiang) published a technical report on Mach-Mind-4-Flash, a Mixture-of-Experts (MoE) model with 35 billion total parameters and roughly 3 billion activated per token. The central claim is that through post-training optimization alone, without additional pretraining compute, the model reaches performance on par with or above 100-billion-parameter-class systems.

The method, as described in the report, is a unified RL pipeline that orchestrates multiple expert tracks, a multi-teacher distillation stage, and a dedicated token-efficiency stage, plus scalable agentic interaction environments for large-scale reinforcement learning aimed at real-world application tasks.

The framing matters as much as the numbers. An automaker's AI group, not a dedicated frontier lab, is arguing that the marginal capability now comes from post-training and environment design rather than from raw scale. That claim awaits independent reproduction on named benchmarks, and "100B-class" is not a fixed standard, depending heavily on which baseline and which tasks are chosen. What is concrete is another downloadable Chinese model built to achieve frontier-adjacent behavior with a small number of activated parameters.

Veracity: Plausible

54/100

If true, who benefits: China's open-weight ecosystem, compute-constrained governments and enterprises that want cheap-to-serve models, and Li Auto's standing as an AI developer, at the expense of compute-rich pretraining incumbents.
The nuance: "100B-class" is not a fixed standard and rests on self-reported benchmarks, and the model is initialized from Qwen3.5-35B, so the gains are post-training on an existing base rather than a result built from scratch.

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

If small activated-parameter MoE models can be raised to larger-model behavior with post-training alone, the advantage shifts from compute-rich pretraining firms toward teams with strong RL environment design and distillation pipelines, and the beneficiaries are deployers who want cheap-to-serve models. It also strengthens the non-US open-weight ecosystem, since another capable Chinese model with a small serving cost is exactly what compute-constrained governments and enterprises standardize on.

What to watch

Independent evaluations of Mach-Mind-4-Flash on public agentic and reasoning benchmarks against a stated 100B baseline, which will confirm or undermine the parity claim.
Whether more non-lab industrial firms publish frontier-adjacent models, signaling that post-training know-how is diffusing beyond the dedicated labs.

Observations to monitor, not financial advice.

2 sources

Synthesized from: Polylog editors · arXiv

Part of a tracked trend

Frontier Model Efficiency Gains

Capability per unit of training and inference compute keeps improving, letting newer models match prior frontier performance far more cheaply and gradually loosening the link between raw scale and capability.