Particle-1 cut output tokens by 28 to 52 percent against DeepSeek V4.1 Flash while holding pass rates, and the savings came from thinking less
AspireAI community · Particle
Particle, a new company from Mixpanel founder Suhail Doshi, released Particle-1, a model post-trained on DeepSeek V4.1 Flash and trained in internally built reinforcement-learning environments with a reward function that pays for solving the task with fewer tokens. Across four coding benchmarks (DeepSWE 1.1, Terminal-Bench 2.1, ReactBench, ADE Bench) it uses 28.4 to 52.2 percent fewer output tokens and 34.4 to 51.6 percent lower cost per task than the base model, with pass@1 inside the base model's confidence intervals; mean cost per task was $0.2211 against $0.3368 for DeepSeek V4.1 Flash and $0.6523 for GLM 5.3 Flash. Trajectory analysis attributes roughly 90 percent of the savings on three benchmarks to reasoning tokens rather than tool-call text: the model deliberates less for the same actions and stops once a task verifies. It is post-trained with a rank-64 LoRA adapter covering the experts in all layers, about 1.25 percent of parameters. The honest caveat is in their own numbers: scaling the adapter to 8x cut output a further 78.7 percent but left only three of eight legal reports passing the strict check, so aggressive efficiency pressure compresses useful qualifications along with redundant analysis.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.