Perplexity Open-Sources Decider v1.2: Release and Early Benchmarks

Episode 1 · Perplexity Open-Sources Decider v1.2 Decision Models in 27B and 4B (2026-10-17, 17 posts)

Perplexity team member antoinechaffin has released a new generation of open-weight multimodal decision models: pplx-decider-v1.2 (27B) and pplx-decider-flash-v1.2 (4B), with improved results on both text and vision tasks and inference speedups of up to 5x. The models are available on Hugging Face under the Apache license and accessible via the Perplexity API, with further price cuts for the flash model. The release is notable for its transparent results disclosure and the strong performance of the distilled small model.

Confirmed

• Two sizes released: pplx-decider-v1.2 (27B) and pplx-decider-flash-v1.2 (4B), both open-weight multimodal decision models trained with substantially more compute and data
• The 27B scores 67.7 on the public split, 64.0 on same-skill, and 65.5 on Full, ranking 3rd on the Decision Index with 144ms median latency, behind the newly crowned leader StarLux-4B; scores on some new domains actually regressed versus the previous generation
• The 4B version shares all design decisions with the 27B (including the attention schema and training recipe); the only difference is that it learns from signals distilled from the 27B model, performs exceptionally well for its size, and tops the 4B tier
• The models are public on Hugging Face under the Apache license; also available via the Perplexity API, with further price reductions for the flash model
• The author addressed questions about public/private benchmark gaps, explaining how the private set differs from the public one

Why it matters

• Open weights plus the Apache license lower the barrier to using decision models, and the distillation approach demonstrates the potential of small models
• The author proactively disclosed actual rankings and weaknesses (regressed new-domain scores, 3rd place rather than 1st) and responded to benchmark criticism, making the disclosure relatively transparent

• Perplexity open-sources next-gen decider models in 27B and 4B sizes with up to 5x faster inference — antoine_chaffin · 2026-10-17
• pplx-decider packs multiple questions per state into one bidirectional forward pass — antoine_chaffin · 2026-10-17
• Packing multiple questions into one forward pass saves compute but adds instability — antoine_chaffin · 2026-10-17
• Perplexity says rivals train on Decision Index benchmark sets, opts to follow for fair comparison — antoine_chaffin · 2026-10-17
• Perplexity details decider data hygiene: official splits only, deduped, 16.6% of one stage — antoine_chaffin · 2026-10-17
• Perplexity traces decider v1.2 regressions to longer training, mitigates via model souping — antoine_chaffin · 2026-10-17
• Perplexity's 4B decider model distills from its 27B sibling using the same training recipe — antoine_chaffin · 2026-10-17
• Perplexity decider v1.2 lands #3 on Decision Index as longer training erodes backbone skills — antoine_chaffin · 2026-10-17
• pplx-decider models open-sourced under Apache license on Hugging Face — antoine_chaffin · 2026-10-17
• Perplexity ships open-weights pplx-decider-v1.2 with up to 5x faster inference — inductionheads · 2026-10-17
• Perplexity open-sources Decision Models: 4B at 1 cent/M tokens, tops HF Decision Index — antoine_chaffin · 2026-10-17
• Perplexity's open-weights Decider v1.2 (27B) and Flash v1.2 (4B) hit OpenRouter — bo_wangbo · 2026-10-17
• Perplexity's open-weights Decider v1.2 (27B) and Flash v1.2 (4B) land on OpenRouter — antoine_chaffin · 2026-10-17
• Perplexity open-sources pplx-decider v1.2 multimodal decision models, Flash up to 5x faster — denisyarats · 2026-10-17
• Perplexity's Decider V1.2 decision models land on Vercel AI Gateway — AravSrinivas · 2026-10-17
• Perplexity's pplx-decider tops internal 72-task reranker bench with zero reranking training — bo_wangbo · 2026-10-17
• Perplexity reveals pplx-decider: reranker with zero reranking supervision, private MTEB-style eval — bo_wangbo · 2026-10-17

Episode 2 · Perplexity's Decider Flash Matches Wald Accuracy at Nearly Double Speed (2026-10-17, 2 posts)

Third-party tests show Perplexity's local model pplx-decider-flash-v1.2 matches the accuracy of benchmark leader Wald while running nearly twice as fast on a 16GB M4 Mac mini.

• Perplexity Flash matches Wald's accuracy at nearly 2x speed in local agent-decision benchmark — WolframRvnwlf · 2026-10-17
• Perplexity Decider Flash matches Wald accuracy at nearly 2x speed on M4 Mac — antoine_chaffin · 2026-10-18