{"id":"reax-1-3","slug":"reax-1-3","name":"REAX 1.3","orgId":"reax","family":"reax-1-3","releaseDate":null,"availability":"ga","openWeights":true,"license":"Apache 2.0 for the blended work, with upstream licenses and notices preserved (component weights Apache 2.0 or MIT); REAX Engine inference code MIT","paramsTotal":57358419036,"paramsSource":"huggingface","contextWindow":null,"knowledgeCutoff":null,"modelCardUrl":"https://x.com/reaxlabs/status/2117664350655729995","hfRepo":"decisionmodels/REAX-1.3","modalities":["text"],"aliases":["REAX 1.3","reax-1.3","decisionmodels/reax-1.3","reax13","reax v1.3","reax 1.3"],"aliasNote":null,"origin":"tracked","tracked":{"trackedSince":"2026-11-03","announced":"2026-11-03","blurb":"REAX's open decision model: two weight-blended parts (one Gemma 4 31B lineage, one Qwen3.5-27B lineage) run in parallel on one GPU and their per-option probabilities are averaged 50/50. Each part blends a model from REAX's Bittensor subnet 92 miner competition with an open decision model. Open weights (Apache 2.0) on Hugging Face; hosted as reax-1.3 on the Decision Models API.","api":"live","apiModelId":"reax-1.3","openWeights":"yes","params":{"text":"About 57B across two parts: part G about 31B (62.5 GB in bf16, Gemma 4 31B lineage) and part Q 26.1B (26,085,330,160, Qwen3.5-27B lineage)","source":"https://huggingface.co/decisionmodels/REAX-1.3"},"facts":[{"label":"Type","value":"Decision model: returns a probability for every supplied option through a Jev-compatible interface (POST /v1/systemone)"},{"label":"Architecture","value":"Two parts run on every request and their per-option probabilities are averaged 50/50"},{"label":"Part G","value":"Weight-space blend of CharlieBT/gx31-decider (subnet 92 miner uid 18; Gemma 4 31B lineage via tattrongvu/Jeffan-31b-1710) and PelaAI/KnowLine-31B-Gen1"},{"label":"Part Q","value":"Weight-space blend of gurujustin/sn92-gen-p5 (subnet 92 miner uid 83; Qwen3.5-27B lineage via baselabs/mammoth-1) and vllm-sr/d3-ultra"},{"label":"Blending","value":"No further fine-tuning after blending; blend proportions not disclosed"},{"label":"Hardware","value":"Both parts in FP8 on one NVIDIA RTX PRO 6000 Blackwell 96 GB (about 88.5 GiB peak reserved); about 115 GB download"},{"label":"Context","value":"Requests above 32,000 input tokens are refused, not truncated (open engine); the hosted API takes up to 8,192 input tokens and one question per request"},{"label":"Weights","value":"Hugging Face decisionmodels/REAX-1.3, not gated; served with the open-source REAX Engine (reax13 mode)"},{"label":"License","value":"Apache 2.0 for the blended work, with upstream licenses and notices preserved (component weights Apache 2.0 or MIT); REAX Engine inference code MIT"},{"label":"API","value":"Decision Models API (decisionmodels.io) model reax-1.3, EU tier only, text only, EUR 0.40 per 1M input tokens, output free"}],"claims":["Lab claim: 73.71 on the full public Decision Index 0.3 suite (REAX's own run, all 140,178 scoreable requests, FP8 engine); REAX says this is not an official leaderboard score.","Lab claim: 69.5 on 3,478 public Hugging Face Decision Index items against 68.8 for d3-ultra (within noise) and 67.4 for KnowLine-31B-Gen1, in a self-run comparison REAX calls a manual selection.","Lab claim: 207 ms median latency (655 ms mean, 741 ms p80) on one RTX PRO 6000, one request at a time."],"sources":[{"kind":"lab","title":"REAX announcement on X","url":"https://x.com/reaxlabs/status/2117664350655729995","date":"2026-11-03"},{"kind":"lab","title":"Decision Models model lineup (REAX 1.3 API listing)","url":"https://decisionmodels.io"},{"kind":"repo","title":"decisionmodels/REAX-1.3 on Hugging Face","url":"https://huggingface.co/decisionmodels/REAX-1.3"}],"firstBenchmark":{"date":"2026-11-06","benchmarkId":"decision-index-full"}},"decision":{"kind":"decision","outputTypes":["noul","choice","score","multi"],"vision":false,"backbone":null,"technique":"autoregressive","weightsRepo":null,"servedParams":57358419036,"hosted":null,"coverage":1,"partialRun":false,"latencyMedianMs":245.9,"latencyHosted":false,"firstSeen":"2026-11-06","reliability":[{"w":0.0009,"acc":0.0435,"conf":0.0884},{"w":0.0125,"acc":0.1285,"conf":0.153},{"w":0.0113,"acc":0.2473,"conf":0.2474},{"w":0.0243,"acc":0.3406,"conf":0.3581},{"w":0.0383,"acc":0.4392,"conf":0.4561},{"w":0.0813,"acc":0.5404,"conf":0.5479},{"w":0.0769,"acc":0.6059,"conf":0.6484},{"w":0.0757,"acc":0.6687,"conf":0.7496},{"w":0.0903,"acc":0.8396,"conf":0.8542},{"w":0.5884,"acc":0.9736,"conf":0.9824}],"priceUsdPer1k":null,"api":false,"measuredOn":null},"organization":{"id":"reax","name":"REAX","slug":"reax","country":null,"url":"https://decisionmodels.io","color":"#11110F","mark":"R"},"price":null,"answer":{"word":"Close.","scope":"Near-SOTA in Decision models, rank #3.","lead":"As of Nov 22, 2026, REAX 1.3 is near state of the art, rank #3 in Decision models (since Nov 22, 2026; sources Decision Index Full; methodology v1.0). Close in Decision models: rank #3, 0.9 percentile points behind the leader.","confidence":"medium"},"status":"BENCHMARKED","verdicts":[{"id":"v_overall_decision_reax-1-3_2026-11-22_open_NEAR_SOTA","model":"reax-1-3","model_name":"REAX 1.3","category":"decision","category_name":"Decision models","slice":"overall","state":"NEAR_SOTA","rank":3,"present_evidence":[],"missing_evidence":[],"coverage":null,"segments":[],"confirm":[],"margin":0.8695652173912976,"confidence":"medium","valid_from":"2026-11-22","valid_to":null,"configuration":null,"rationale":"Close in Decision models: rank #3, 0.9 percentile points behind the leader.","safeguard_flags":[],"co_sota_with":[],"challengers":[],"methodology":"1.0","label":"Near-SOTA · Decision models"},{"id":"v_open-weights_decision_reax-1-3_2026-11-22_open_NEAR_SOTA","model":"reax-1-3","model_name":"REAX 1.3","category":"decision","category_name":"Decision models","slice":"open-weights","state":"NEAR_SOTA","rank":3,"present_evidence":[],"missing_evidence":[],"coverage":null,"segments":[],"confirm":[],"margin":0.9463722397476317,"confidence":"medium","valid_from":"2026-11-22","valid_to":null,"configuration":null,"rationale":"Close in Decision models: rank #3, 0.9 percentile points behind the leader.","safeguard_flags":[],"co_sota_with":[],"challengers":[],"methodology":"1.0","label":"Near-SOTA · Decision models (open weights)"},{"id":"v_calibration_decision_reax-1-3_2026-11-22_open_RANKED","model":"reax-1-3","model_name":"REAX 1.3","category":"decision","category_name":"Decision models","slice":"calibration","state":"RANKED","rank":6,"present_evidence":[],"missing_evidence":[],"coverage":null,"segments":[],"confirm":[],"margin":0.0043,"confidence":"medium","valid_from":"2026-11-22","valid_to":null,"configuration":null,"rationale":"Ranked #6 in Decision models calibration. REAX 1.3 is inside the top quartile of accuracy.","safeguard_flags":[],"co_sota_with":[],"challengers":[],"methodology":"1.0","label":"#6 in Decision models (calibration)"}]}