Is XERJ Query-1 state of the art?Pending benchmarks.Announced Nov 5, 2026; no independent results yet.
XERJ's proprietary neurosymbolic query model compiles natural-language questions into short executable search programs over user-written typed facts, rather than generating prose answers. The engine executes and verifies the programs, returning ANSWER, UNKNOWN or CONFLICT with evidence traces. Public model bundle available with rc.101 on GitHub releases; local CPU serving is opt-in via --facts-model-dir. The engine is Apache-2.0; the model and its training/serving components are not open. No public hosted API stated.
The details so far
- Lab
- XERJ
- Announced
- Nov 5, 2026
- API
- Not yet
- Open weights
- Closed
- Role
- Neurosymbolic question-to-program model over typed facts; not a general-purpose LLM
- Programs
- LOOKUP, CHAIN, comparisons (LT/GT/EQ/NOT), PICK_LT/PICK_GT and set-count programs (SET/FIND/FILTER/AGG COUNT); short programs executed by the engine
- Outputs
- ANSWER, UNKNOWN or CONFLICT; program and record/version/source-span evidence traces
- Access
- Public GitHub rc.101 model-bundle download, no signup required; start xerj with --facts-model-dir to enable (off by default)
- License
- Proprietary model and training/serving components; Apache-2.0 engine. Public download does not mean open weights
- Bundle
- Separate model-bundle release asset (408,024,623-byte compressed archive); installer describes about 460 MiB
- Serving
- Local CPU; int8 with calibrated guard/fallback path. No public hosted API stated
- Input limits
- Users or agents must write typed facts; no automatic fact extraction from arbitrary documents. Counts require stable subject_id
- Operation limits
- Set counts only, not sums, averages, grouped counts or ratios; no joins beyond one hop; programs beyond six steps or four stack slots return UNKNOWN
- Known limitations
- Lab discloses over-declining, unmeasured fresh-wording robustness for this bundle, and a count-noun verification defect: a member count can be returned when the question asks for an unsupported quantity. Large namespaces can be slow and return UNKNOWN
- Related
- XERJ Query-1 is the named model, separate from the XERJ search engine and its rc.101 software release
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: about 120 ms median ask latency in int8-guard serving mode on a quiet x86 machine with four CPU cores and a namespace of up to 8,000 facts; not the separate shared-aarch64 head-to-head timing.
Lab claim: on its 160-question head-to-head over pre-extracted facts, Query-1 returned 124 right answers, 1 wrong and 35 declines. Questions came from its own evaluation distribution; these are lab results, not independent board evidence.
Sources
LABXERJ Query-1 announcement (XERJ Engineering)xerj.org/blog/xerj-query-ai-model · Nov 5, 2026LABXERJ Query-1 model pagexerj.org/xerj-query-1REPOXERJ rc.101 release and public model bundlegithub.com/xerj-org/xerj/releases/tag/v1.0.0-rc.101 · Nov 5, 2026REPOXERJ engine repositorygithub.com/xerj-org/xerjLABXERJ founder on X: Query-1x.com/d0znpp/status/2118840938147770450 · Nov 6, 2026Tracking ends Feb 4, 2027 unless the API or weights become available.