Is hs-decide-1-flash state of the art?Pending benchmarks.Announced Nov 3, 2026; no independent results yet.
Vectorize's 0.8B decision model for the Hindsight agent-memory system: instead of generating text, it returns a probability for every answer option in one forward pass. Open weights (Apache 2.0) on Hugging Face, also served on the Hindsight Inference API.
The details so far
- Lab
- Vectorize
- Announced
- Nov 3, 2026
- API
- Live · hs-decide-1-flash · Hindsight Inference (inference.hindsight.vectorize.io) POST /v1/systemone, System One compatible, $0.02 per 1M input tokens; also accepted on the Cohere-compatible /v1/rerank to drop irrelevant documents
- Open weights
- Published · Hugging Face hindsight-ai/hs-decide-1-flash, bfloat16, not gated; serves on vLLM with a plugin that adds POST /v1/systemone
- Parameters
- 0.8B (752,393,024 stored parameters in the safetensors)
- Type
- Decision model: one forward pass returns a probability for each answer option; it generates no text
- Question types
- choice (2 to 255 options), score (2 to 10 ordered levels) and noul (yes or no); several questions can share one request
- Base model
- Qwen/Qwen3.5-0.8B
- Context
- Trained on requests up to 65,536 tokens; longer requests are rejected
- Languages
- English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese
- License
- Apache 2.0
- Use in Hindsight
- Opt-in in Hindsight 1.2.0 for consolidation, reflect and knowledge-page judgments, plus optional recall pruning and dedup
- Siblings
- hs-embed-1 and hs-rerank-1, released the same day (separate pages)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: inside Hindsight, six of seven LLMs give better answers and five get cheaper with hs-decide-1-flash taking the small judgments; on gemini-3.8-flash the operations it joins cost 2.5x less, spend 1.9x less time in model calls and score 2.2 points higher, while reflect quality dips on some LLMs.
Lab claim: median 167 ms per request (350 ms p90).
Lab claim: mean of 75.6 over eight decision benchmarks against 74.5 for StartLux-Decision 0.8B Instant, 73.4 for d3-lite and 66.6 for decider-0.8b, though it trails on the public Decision Index (36.1 against 53.8 for StartLux).
Sources
LABIntroducing Hindsight Models (Hindsight blog)hindsight.vectorize.io/blog/2026/11/03/hindsight-models · Nov 3, 2026LABVectorize announcement on Xx.com/Vectorizeio/status/2117647994048704588 · Nov 3, 2026LABhs-decide-1-flash benchmarks (Hindsight Benchmarks)benchmarks.hindsight.vectorize.io/models/decisionREPOhindsight-ai/hs-decide-1-flash on Hugging Facehuggingface.co/hindsight-ai/hs-decide-1-flashTracking continues until an independent board lists it.