Perplexity has published pplx-decider-v1.1-27b, a 27-billion-parameter model that picks one answer from a menu, such as which team should handle a ticket, and it credits most of a 5.16-point jump to removing a restriction nearly every language model carries. Cloudflare’s Clef-omni and Microsoft-Decision-1 arrived in the same news cycle, but this card is the only one of the three that says what changed inside the model.
The numbers come from Perplexity’s own Hugging Face model card. The Decision Index score rose from 56.4 for the first version to 61.56. On the same table, Jev scores 57.9, so the card’s “more than 3.5 points” lead is 3.66 points by plain subtraction. The card says the new model uses the same backbone as before and names it as Qwen3.8-27B.
The restriction is the causal mask. A normal language model reads left to right: each token may look only at the tokens before it, because the model is trained to predict the next word and must not see the answer. A model that sorts a support request into billing, support, or sales has no next word to protect. The whole request is already in front of it, so every word can be read in light of the words that follow. A qualifier at the end of a sentence can reverse the meaning of the first clause, and under a causal mask the early tokens never see it.
Perplexity says most of the gain came from lifting the mask and from training on more data, including tasksource, a public collection of language tasks. The card does not split the credit, so nobody outside Perplexity can say how many of the 5.16 points belong to the mask. It also qualifies the change. The change applies to the full-attention layers alone. The remaining layers, which use a cheaper linear form of attention, keep their original behavior, so the model reads the whole input only in part of its stack.
That detail carries a serving cost. According to the card, ordinary causal inference will not reproduce the published results, and the output is a 255-way decision head rather than a vocabulary head. A team that wants to run it inside a standard text-generation stack needs a custom export that preserves the attention mode and applies the saved calibration temperature exactly once.
The Decision Index is the weaker part of the story. The card lists five categories, namely knowledge, language, retrieval, tools, and arts, and says the overall figure uses “the suite’s weighting.” It does not say who maintains the index, whether anyone outside Perplexity has run it, or who built Jev. Perplexity chose the benchmark and the rival, and the table is the company’s own. The precision is uneven too: the new model gets two decimals while Jev and the old version get one, which suggests the columns were not produced in a single pass.
The category rows are more informative than the headline. Jev still wins knowledge, 51.4 to 48.18. Tools slipped slightly from the first version, 79.3 to 78.88. The biggest gains landed in knowledge at 7.28 points, retrieval at 6.36, and language at 5.95, which are the categories that depend most on reading a passage in full.
What Perplexity actually shipped is murky. The card describes about 49 GiB of weights, a CUDA GPU, and “authenticated Hugging Face access to this private repository.” It names no licence and says nothing about an API. Calling this open would be a guess, and a company planning to deploy it should assume permission has not been granted until the licence appears. Images are supported, with an 8,192-token limit per decision, and the reported evaluation capped them at 512 by 512 pixels.
With three vendors selling the same kind of model, the comparison that matters is cost per correct decision, and the card publishes no latency, throughput, or price. A 27B model competes with smaller, cheaper classifiers on that measure, not on intelligence. A team choosing among the three should run a few hundred of its own labelled requests through each, and confirm that the Perplexity checkpoint runs noncausal, because a run in the default mode will score below the card.
Perplexity, via the pplx-decider-v1.1-27b model card on Hugging Face, which carries no publication date.
