Perplexity pplx-decider: Open 27B Decision Model Forked to NVFP4

A model that picks, not writes

On October 1, 2026, Perplexity launched a Decisions API powered by pplx-decider-v1-27b, an open-weight model fine-tuned from Qwen3.8-27B and published on Hugging Face under the Apache 2.0 license 3. The model does not generate free-form text. It takes a set of choices and returns probabilities, with no written answer attached 2. The API supports three typed response shapes: yes/no, multiple choice, and ordered score 3.

The pricing is aggressive. Input costs $0.04 per million tokens, and output is free 23. Perplexity's documentation lists latency under two seconds for a few hundred input tokens, rising to about 23 seconds near the model's 262,144-token context limit 3. Running the model yourself is a heavier commitment. The full BF16 weights need roughly 49 GiB of memory 2.

The benchmark claims, and the pushback

Perplexity's headline number comes from its own panel of 11 benchmarks covering 7,210 samples. On that panel, pplx-decider scored 85.71%, rival Jev scored 84.51%, and the base Qwen model scored 74.76% 23. The largest margin was on RAGTruth, where Perplexity's model reached 88.80% to Jev's 77.27% 3.

The aggregate lead looks less decisive up close. Jev still wins 6 of the 11 individual tests, and the panel was assembled by Perplexity itself 2. Some responses to the announcement were skeptical. One commenter affiliated with Fastino Labs accused Perplexity of cherry-picking benchmarks and pointed to a separate comparison of Jev, GLiDE, and Perplexity. Another user said a model called Clef beat Perplexity in their own tests, and others asked what the benchmark mix actually was 2.

Taken together, this supports a cautious reading. The roughly 11-point gain over base Qwen suggests the fine-tune does real work. The 1.2-point edge over Jev, however, rests on a vendor-chosen panel that is driven largely by one outsized result. The comparison worth watching is how the model performs on independent evaluations.

The community fork arrived almost immediately

The permissive license had a visible effect within days. AveLabs published Qwen3.8-27B-Perplexing-Pegasus, a repackaging of the decider in three formats: BF16 weights, GGUF quantizations for consumer GPUs, and an NVFP4 build aimed at NVIDIA's Blackwell hardware 1.

The build process is documented in detail. AveLabs loaded the pplx-decider backbone, vision tower, and readout layer. It then confirmed that each readout row matched the base Qwen3.8-27B output-head row for its corresponding answer-code token, with cosine similarity of 1.0000 1. This indicates that Perplexity's decision readout is effectively a slice of the original language-model head. AveLabs restored the full lm_head and 15 multi-token-prediction tensors from the upstream Qwen release at a pinned revision. That restoration brings back general chat ability alongside the decision behavior 1.

Quantization used NVIDIA's ModelOpt with a mixed scheme [1]:

• MLP weights and activations in NVFP4
• Attention and the KV cache in FP8
• Output and MTP heads kept in BF16
• Calibration on 1,024 sequences of 512 tokens from cnn_dailymail

The repository says decisions were validated against the unmodified original through identical requests to the systemone endpoint. Chat quality was checked against stock Qwen3.8-27B NVFP4 1.

Why it matters

This release shows a pattern in open-weight AI. A company ships a narrow, commercially useful model under a permissive license, and the community reshapes it for new hardware and use cases almost immediately. Perplexity's pitch is cheap, structured judgments for classification, routing, and fact-checking pipelines, an area where generative output is often unnecessary. Free output tokens fit that design, since the answer is a probability distribution rather than prose.

The AveLabs fork matters for two reasons. First, it lowers the hardware barrier. The 49 GiB BF16 footprint 2 rules out many setups, while NVFP4 and GGUF variants open the model to Blackwell and consumer GPUs. Second, AveLabs' readout check offers a useful look at how Perplexity built the decider. It appears to be a fine-tuned backbone reading out over a fixed set of answer tokens, not an architecturally new system 1.

The open questions are on evaluation, not engineering. Perplexity's own panel favors its model by a narrow aggregate margin, and competitors are already disputing the framing 2. Openly available weights will make independent testing easy. That testing, more than vendor benchmarks, will decide whether pplx-decider becomes a default tool for automated decisions.