I-Cisco Foundation AI Ikhipha i-Antares: 350M kanye ne-1B Amamodeli Wesisindo Esivulekile Aveza Ubungozi Obaziwayo Ngaphakathi Kwamakhodi Angempela
I-Cisco Foundation AI ikhishwe Antaresumndeni wokuphepha wamamodeli wezilimi ezincane (ama-SLM) akhelwe umsebenzi owodwa omncane wokuvikela. Umsebenzi ukwenza kwasendaweni ukuba sengozini. Uma unikezwe incazelo yokuba sengozini kanye nenqolobane, thola amafayela aqukethe iphutha.
Amamodeli amabili anesisindo esivulekile futhi ayatholakala manje ku-Hugging Face, Antares-350M kanye ne-Antares-1B. Zombili ziyi-Apache 2.0. Ithimba le-Cisco liphinde lathumela i-Vulnerability Localization Benchmark (Ibhentshi le-VLoc), ukuhlolwa kwe-ejenti yemisebenzi engu-500, ngaphansi kwelayisensi efanayo.
Umphumela oyinhloko awusona isimo esisha sobuciko. Kungenxa yokuthi imodeli ye-1B ifinyelela ku-0.209 File F1. I-GPT-5.5 ifinyelela ku-0.229, futhi imodeli yesisindo esivulekile engu-753B ifinyelela ku-0.186.
Inkinga i-Antares ibhekelwa kuyo
Ukuvikeleka kwesofthiwe kuncike ekuxhumeni ulwazi lwangaphandle lokuba sengozini kwikhodi yomthombo yangaphakathi. Lolo lwazi luhlala kusizindalwazi somphakathi, izeluleko, kanye nokubalwa kobuthakathaka obujwayelekile. Ikhodi ihlala kumakhosombe amakhulu, amamojula, futhi anothile ngokuncika.
Ukuxhuma kokubili kuyabiza. Ama-Devs asesha ikhodi engaziwa, alandele izimiso zokuqamba amagama, ahlole izindlela zekholi, futhi aqhathanise amafayela ekhandidethi. Uhlaka lweCisco ukuthi lesi sinyathelo sokuqala sokunquma yilapho izindleko zigxila khona.
I-Antares ayithathi indawo yochungechunge lokuphepha lohlelo lokusebenza. UCisco ubeka ingcaca ngalokhu. Amathimba we-Dev asadinga ukuskena kokuncika, ukuskena okuyimfihlo, ukuhlola okuguquguqukayo, ukuhlolwa kweziqukathi, imodeli yokusongela, nokubuyekezwa kochwepheshe.
Ukuqonda Amamodeli
I-Antares iqukethe ama-decoder-transformer amathathu kuphela kumapharamitha angu-350M, 1B, kanye no-3B. Zontathu ziqala kusuka ezindaweni zokuhlola ze-IBM Granite 4.0. Babelana ngethokheni nezakhiwo: ukunakwa kwemibuzo eqoqiwe, ama-SwiGLU MLPs, i-RMSNorm, i-RoPE, nokushumeka okwabiwe okokufaka/okuphumayo.
| Imodeli | Amapharamitha | Indawo yokuhlola eyisisekelo | Umongo | Izendlalelo / ezifihliwe / amakhanda e-KV | Isimo |
|---|---|---|---|---|---|
| I-Antares-350M | 350M | I-Granite 4.0 350M | 32K | 28/1024/4 | Vula izisindo |
| I-Antares-1B | 1.6B | I-Granite 4.0 1B | 128K | 40/2048/4 | Vula izisindo |
| I-Antares-3B | 3B | I-Granite 4.0 Micro | 128K | Akushicilelwe | Ayikhululwa |
Isebenza kanjani i-Agent Loop
I-Antares ayihlolwa njengemodeli yokulandelana ezimele. Igijima ngaphakathi kweluphu eboshiwe ngamathuluzi amathathu.
Imodeli ithola incazelo yesigaba se-CWE futhi akukho okunye. Awukho umbhalo wokweluleka, awekho amacebo efayela, akukho mininingwane yobunzima. Ibese ikhipha imiyalo yokugcina efundwayo kuphela ngokumelene ne-sandbox ye-Docker enenethiwekhi evaliwe. Umyalo ophumayo uncishiswa ube yizinhlamvu ezingu-2,000 ngaphambi kokufaka okulotshiweyo.
Isabelomali yizingcingo eziyi-15 zomsebenzi ngamunye. Imodeli iphetha ngokushaya ucingo submit_vulnerable_files ngohlu olusezingeni, noma submit_no_vulnerability_found. Ukuhanjiswa ngokwako akubalwa uma kuqhathaniswa nesabelomali.
Okuphumayo kuwuhlu olulinganiselwe lwezindlela zamafayela kanye nokulandelela kokuhlola okukhiqize.
Yiziphi izilinganiso ze-VLoc Bench
I-VLoc Bench idonsa imisebenzi engama-500 kumakhosombe angama-290 ayingqayizivele emhlabeni wangempela. Imithombo iyizeluleko Zokuphepha ze-GitHub ezisesidlangalaleni kuwo wonke ama-ecosystem ayisithupha: npm, pip, Maven, Go, Rust, kanye noMqambi. Ihlanganisa izigaba ze-CWE ezihlukile ezingu-147, futhi u-78% wemingenelo uphethe izihlonzi ze-CVE ezabelwe.
Iqiniso eliyisisekelo lisuselwa kusiqephu sokuphepha. Amafayela ashintshiwe ekulungiseni amalebula, anokuvivinywa, amadokhumenti, nokulungiselelwa kukhishiwe.
Ibhentshimakhi inezigaba ezimbili:
- Isigaba A inikeza imodeli isifinyezo esisengozini kanye nemiphumela Ifayela F1.
- Isigaba B inikeza isifinyezo esinamagatshana kanye nemiphumela Isilinganiso Esibi Sangempela, sihlola ukuthi ingabe imodeli iphakamisa i-alamu engamanga kukhodi egxilile.
Imiphumela: ukuqeqeshwa okuqondene nomsebenzi kudlula isikali sepharamitha
Iphethini kudatha iyiwa lamandla, hhayi ijika lokukala.
I-Antares-3B ifinyelela ku-0.223 Ifayela F1, ngaphansi nje kwe-GPT-5.5 (xhigh) kokuthi 0.229. I-Antares-1B ifinyelela ku-0.209, ngaphezu kwe-GLM-5.2 kumapharamitha angu-753B, okwenza amaphuzu angu-0.186. I-Antares-350M ifinyelela ku-0.135, ngaphezu kwe-Gemma-4-31B ku-0.101 kanye ne-Gemini 2.5 Flash ku-0.102.
I-Antares-1B iphinda irekhode ukukhunjulwa okuphezulu kakhulu kwanoma iyiphi isistimu ehloliwe ku-0.224.
Amathuluzi okuhlaziya aqinile asetshenziswe ngaphansi kokuhlola okufanayo. I-Semgrep ithola amaphuzu angu-0.086 Ifayela F1, i-CodeQL amaphuzu angu-0.023, kanye ne-Horusec amaphuzu angu-0.020. Ukufundwa kweCisco ukuthi izikena ezisekelwe emthethweni zithola amanye amafayela asengozini kodwa azikwazi ukuhlola ngokuguquguqukayo umongo wenqolobane.
Lapho ikhono livela khona
Izindawo zokuhlola eziyisisekelo ze-Granite 4.0 ezingaqeqeshiwe zinemiphumela engu-0.001, 0.000, kanye ne-0.000 yefayela F1 ngaphansi kwephrothokholi efanayo. Banekhono lokubiza amathuluzi futhi basakhiqiza okukhiphayo okonakele ngaphakathi kweluphu ye-agent.
Ukucushwa okugadiwe kwenza ukuphakamisa okusindayo. Iphakamisa izikali ezintathu ziye ku-0.108, 0.188, kanye no-0.198. Ikhophasi ye-SFT ingu-71.5% wokucabanga nge-cybersecurity, u-15.4% wezindlela zokucinga ngekhodi, kanye no-13.1% wocwaningo olujulile nokucabanga okuvamile. Yonke iminonjana yokucabanga ivela kuthisha oyedwa, i-GPT-OSS-120B, ukugwema ukushintsha kokusabalalisa kothisha.
I-GRPO ibe isingeza u-11% kuya ku-25%, ngenzuzo enkulu kakhulu ehlobene ku-350M. Imiklomelo iyaqinisekiswa futhi ibalwe ngokohlelo kusukela kumbhalo we-trajectory, ngaphandle kwemodeli yomvuzo efundiwe. Izingxenye zihlanganisa ikhwalithi yokwenza okwasendaweni, ukuziphatha kokuthunyelwa, ukuthobela ukusetshenziswa kwamathuluzi, ukuhlola, nezijeziso zemiphumela engalungile.
Umphumela wokuhluka ungase ubaluleke ngaphezu kokushiwo. I-GRPO inciphisa ukuchezuka okujwayelekile kokugijima kuya ku-run ngo-42% kuya ku-65%. Ukuqaliswa kokuhlola okukodwa kwe-GRPO kuyisilinganiso esithembeke kakhulu kunomgijimo we-SFT owodwa.
Kukhona futhi ukuhlukaniswa okuncike esikalini kumasu afundiwe. Ngemuva kwe-GRPO, amamodeli we-350M ne-1B asebenzisa imiyalo yokusesha engu-87% kuya ku-89% futhi athumele amafayela engeziwe. Imodeli ye-3B ihlala ekusesheni okungu-52% futhi 37% ifundwe, futhi ithumela amafayela ambalwa ngokunemba okuphezulu. Umklomelo awuzange unqume noma iyiphi inqubomgomo.
Ukuthunyelwa
Okuthathwayo Okubalulekile
- I-Antares-1B ishaya okuthi 0.209 Ifayela F1 ku-VLoc Bench, ngaphezu kwe-GLM-5.2 kumapharamitha angu-753B kanye ne-Gemini 3 Pro.
- Amaphoyinti okuhlola asisekelo e-Granite 4.0 athola amaphuzu angu-~0.000 ngaphansi kwephrothokholi efanayo, ngakho-ke ngemva kokuqeqeshwa kunikeza wonke amandla.
- I-GRPO ingeza u-11-25% wefayela F1 futhi isike ukuhlukahluka kokugijima kuya ku-42-65%, okubaluleke kakhulu kumaskeni we-CI aphindaphindiwe.
- Ukushanela okuphelele kwemisebenzi engu-500 kubiza ngaphansi kuka-$1 ku-H100 eyodwa, uma kuqhathaniswa no-$12.50 we-GLM-5.2 kanye no-$141 we-GPT-5.5.
- Okuhlukile okuqine kakhulu, i-Antares-3B, ayikhululwa, futhi i-Antares ayinazo izinombolo ze-alamu ezingamanga zeSigaba B ezishicilelwe.
Hlola Amamodeli ku-Hugging Face, Benchmark, GitHub Repo kanye Nombiko Wezobuchwepheshe Oshiwo.
U-Michal Sutter uchwepheshe wesayensi yedatha one-Master of Science in Data Science yase-University of Padova. Ngesisekelo esiqinile ekuhlaziyeni izibalo, ukufunda ngomshini, nobunjiniyela bedatha, u-Michal uphuma phambili ekuguquleni amasethi edatha ayinkimbinkimbi abe imininingwane ephathekayo.
