Dokumentirani izidi za lokalne modele HOME, oblačne ponudnike MAL in zgodovinske oddaljene teste SLAIF. Vsak rezultat je prikazan skupaj s testom, ki ga je ustvaril.
Documented results for local HOME models, MAL cloud providers and historical remote SLAIF tests. Every result stays attached to the test that produced it.
GPT-OSS 20B in Gemma4-12B sta kandidata za krmarja; Mellum2.1 je hitra alternativa, Qwen3.8 počasna referenca. Izbor za živo usmerjanje še čaka primerljivi HELM.GPT-OSS 20B and Gemma4-12B are helmsman candidates; Mellum2.1 is a fast alternative and Qwen3.8 a slow reference. Live routing selection awaits comparable HELM evidence.
Kako brati: uspešen HTTP odgovor še ni kakovosten odgovor. Prekinitev zaradi virov ni vsebinski neuspeh. Rezultati treh okolij niso ena skupna lestvica; današnja dosegljivost modelov ni preverjena.
How to read this: an HTTP success is not a quality pass. A resource interruption is not a content failure. These three environments do not form one leaderboard; current model availability has not been verified.
PASS/FAIL določa izvirni ocenjevalnik. DOKONČANO ≠ PASS. Števila v matriki so zapisi po izidih, ne skupna ocena; ponovitve štejejo ločeno. Ni podatka pomeni vrzel tega posnetka, ne neuspeha modela.PASS/FAIL come from the original evaluator. COMPLETE ≠ PASS. Matrix counts are records by outcome, not a combined score; retries remain separate. No data means a gap in this snapshot, not a model failure.
Matriko na telefonu pomakni vodoravno. Za ožji pregled izberi kampanjo.Swipe the matrix horizontally on a phone. Select a campaign for a narrower view.
Osem konfiguracij, sedem osnovnih modelov. V vsakem nizu je po deset nalog na konfiguracijo. Stolpci kažejo vsebinske PASS, ne zgolj HTTP 200; časi so mediane odgovorov v obliki min:s.
Eight configurations, seven base models. Each series has ten tasks per configuration. Bars show substantive PASS, not merely HTTP 200; response medians use min:s.
Nalaganje rezultatov…
Osnovni niz: 32 PASS, 36 FAIL, 12 INCOMPLETE. RHP: 22 PASS, 20 FAIL, 15 INCOMPLETE, 13 prekinitev zaradi virov in 10 zadržanih nalog.
Basic: 32 PASS, 36 FAIL, 12 INCOMPLETE. RHP: 22 PASS, 20 FAIL, 15 INCOMPLETE, 13 resource interruptions and 10 held tasks.
Osem začetnih smoke preverjanj je uspelo. V ločenem HMAL enournem nizu je uspelo 12/13 klicev; Ministral je porabil kontekst 4096 brez končnega odgovora.
Eight initial smoke checks passed. In a separate one-hour HMAL series, 12/13 calls passed; Ministral used the 4096-token context without a final answer.
Izvirni osnovni in RHP niz zgoraj ostaneta nespremenjena (osnovni 32 PASS, 36 FAIL, 12 INCOMPLETE; RHP 22 PASS, 20 FAIL, 15 INCOMPLETE, 13 prekinitev zaradi virov, 10 zadržanih). Enourne ponovitve štirih TIMEOUT in pet dodatnih sklopov so prikazani ločeno in se ne seštevajo z izvirniki. Zadržano (HELD) pomeni, da modela nismo poklicali. HTTP 200 je samo prenos; COMPLETE pomeni zaključen razpored, ne izvedenih vseh generacij. Časi odziva ne vključujejo nalaganja modela, minimumi virov so vzorčeni. Smoke 8/8 in HMAL 12/13 (22. 10.) ostajata ločena zgodovinska povzetka.
The original basic and RHP series above are unchanged (basic 32 PASS, 36 FAIL, 12 INCOMPLETE; RHP 22 PASS, 20 FAIL, 15 INCOMPLETE, 13 resource interruptions, 10 held). One-hour retries of four TIMEOUTs and five additional campaigns are shown separately and are not added to the originals. HELD means the model was not called. HTTP 200 is transport only; COMPLETE means the schedule was finalized, not that every generation ran. Response times exclude model loading; resource minima are sampled. Smoke 8/8 and HMAL 12/13 (22 Oct) remain separate historical summaries.
Dopolnilni podatki (≈0,5 MB) se naložijo enkrat, ko se približaš temu delu strani…Supplementary data (≈0.5 MB) loads once as you approach this part of the page…
MAL navaja deset ponudnikov s profilnimi modeli. Štirje svetovalni izhodi so dokumentirani; trije se po ID-ju ujemajo s trenutno navedenim profilom. To je dokaz izvedenega klica, ne primerljiv test kakovosti T001–T010.
MAL lists ten providers with model profiles. Four advisory outputs are documented; three exactly match a currently listed profile ID. This proves an observed call, not a comparable T001–T010 quality result.
Grok, DeepSeek, GLM
Gemini
ChatGPT, Claude, Qwen, MiniMax, Mistral, Kimi
Tri konfiguracije, trije različni obsegi. Dolžine spodnjih stolpcev so deleži znotraj posameznega testa; med seboj niso splošna lestvica modelov.
Three configurations, three different scopes. Bar lengths show success within each test, not a general ranking across models.
Nalaganje rezultatov…
Zgornji stolpci ostajajo starejši zgodovinski povzetek. Spodaj so preverjeni javni povzetki oddanih tekov: en tek je en model na eni nalogi. PASS določa izvirni zamrznjeni evaluator; PARTIAL zajema tehnične vrzeli, nepopolne odgovore, pisne zasnove (DESIGN_ONLY) in raziskovalne pilote politike z determinističnim reševalnikom. HTTP 200 je samo prenos. Dostopnost velja za posamezen datirani tek, ne za današnji dostop. Strogi nizi, ocene zasnov in piloti niso združeni v eno oceno; to ni lestvica modelov.
The bars above remain the earlier historical summary. Below are validated public summaries of submitted runs: one run is one model on one task. PASS is set by the original frozen evaluator; PARTIAL covers technical gaps, incomplete answers, written designs (DESIGN_ONLY) and research policy pilots with a deterministic solver. HTTP 200 is transport only. Availability applies to each dated run, not to access today. Strict suites, design scores and pilots are not merged into one score; this is not a model ranking.
Nalaganje oddanih tekov…Loading submitted runs…
Poziv OK: 12/13 PASS. Ministral je izčrpal kontekst 4096 brez odgovora; ni dosegel enourne meje. Čas je native HTTP.OK prompt: 12/13 PASS. Ministral exhausted context 4096 without an answer, not the one-hour limit. Time is native HTTP.
GPT-OSS: en predlog, vsebina in diff FAIL. Terra: mehanski pregled uspešen, semantični pregled čaka.GPT-OSS: one proposal, content and diff FAIL. Terra: mechanical checks passed; semantic review pending.
Qwen3.5-9B 6/10 PASS, Gemma4-12B 9/10 PASS. Kontekst 32768, native thinking. Proračun se razlikuje od starejših testov.Qwen3.5-9B 6/10 PASS, Gemma4-12B 9/10 PASS. Context 32768, native thinking. Budgets differ from older tests.
Izvedene celice RHP. Delni posnetek ni dokaz zaključene celotne baterije.Executed RHP cells. A partial snapshot does not prove completion of the full battery.
Naročninski CLI, xhigh: osnovni testi, RHP in HMAL OK. HELM in L1 sta v namenskih razdelkih. Čas vključuje CLI režijo.Subscription CLI, xhigh: baseline, RHP and HMAL OK. HELM and L1 have their own sections. Time includes CLI overhead.
Ocenjevanje v teku, 10/21 veljavnih. Ocenjevalci so slepi; odgovori, imena modelov in ocene bodo objavljeni šele po objavi seštevka.Judging in progress, 10/21 valid. Judges are blind; answers, model names and scores will be published only after the tally is released.
Lokalni HOME32 je bil 27. 10. pripravljen in zadržan, brez inference. Izvedene Terra celice so ločene; izbor krmarja čaka HELM-R1.1.Local HOME32 was prepared and held on 27 Oct without inference. Executed Terra cells stay separate; selection awaits HELM-R1.1.
Posnetek podatkov: 28. 10. 2026. Časi HOME so časi odgovorov; SLAIF znani subtotal izbranih korakov; HMAL in lahki modeli native HTTP; Terra vključuje CLI režijo. Te meritve nimajo enakega proračuna. Zbrano in mehansko veljavno še ne pomeni semantičnega PASS.Data snapshot: 28 Oct 2026. HOME timings are response times; SLAIF is the known selected-step subtotal; HMAL and light models use native HTTP; Terra includes CLI overhead. Budgets differ. Collected and mechanically valid does not mean semantic PASS.
HOME kakovostni PASS zahteva pravilen format, shemo in vsebino. HTTP 200 pomeni samo uspešen prenos. Osnovni testi so tekli po protokolu cold_per_test; neprekinjena serija na naloženem modelu ni dokazana. Cloud MAL nima potrjenega enotnega T001–T010 teka. SLAIF izidi pokrivajo različne naloge in zgodovinski dostop.
HOME quality PASS requires valid format, schema and content. HTTP 200 only confirms transport. The basic tests used cold_per_test; a continuously loaded series is unproven. Cloud MAL has no verified uniform T001–T010 run. SLAIF outcomes cover different tasks and historical access.
Spodnji naslovi določajo izvirna dokazila. Zasebni Drive ID-ji, končne točke in poverilnice niso vključeni v javno stran.
The titles below identify the source reports. Private Drive IDs, endpoints and credentials are not included in this public page.