LFM2.5-24B-A2B Uyasebenza

I-LFM2.5-24B-A2B iyimodeli yokucabanga enezisindo ezivulekile ekhishwe yi-Liquid AI, futhi iyimodeli ehamba phambili ochungechungeni lwe-LFM2.5. Inamapharamitha angu-24B isiyonke, kodwa ivula angu-2.45B kuphela ngethokheni ngayinye, inesakhiwo se-MoE esinezendlalelo ezingu-40, futhi isekela umongo ongu-131K nezilimi ezingu-16. I-LFM2.5-24B-A2B ikhiqiza uchungechunge lokucabanga ngaphambi kokuphendula, futhi yenzelwe ngokukhethekile ama-Agent asebenza kudivayisi. Ithole u-48.09 ku-LiveClaw, yadlula i-Qwen3.6-35B-A3B enamapharamitha amaningi ngo-50%; ku-M5 Max ifinyelela ku-184 tok/s, okuyijubane eliphindwe ngo-1.7 kune-MoE eshesha kakhulu esigabeni sayo. Ifanele izimo zokucabanga nokubiza amathuluzi lapho kusetshenziswa ukuthunyelwa kwasendaweni okuyimfihlo; emisebenzini edinga ulwazi oluningi kunconywa ukuyihlanganisa nokusesha kolwazi.

Ukucabanga kwe-Agent kudivayisi: Yakhelwe ama-Agent ashesha futhi ayimfihlo asebenza endaweni kumadivayisi omuntu.
Imodi yokucabanga kuphela: Ikhiqiza ngokuzenzakalelayo uchungechunge lokucabanga olusobala ngaphambi kokuphendula, isebenzisa izindleko eziphansi zokuvula ukuze ithuthukise ikhwalithi yezimpendulo.
Ukubiza imisebenzi: Isekela ngokwendabuko ukugeleza kwezigaba ezine kokubiza amathuluzi, futhi ingaxhunywa ngokuqondile ku-OpenClaw, Hermes Agent, Pi, OpenCode nezinye izinhlaka.
Ukuthunyelwa ngamafomethi amaningi: Ihlinzeka ngezinguqulo zomdabu, FP8, GGUF, ONNX ne-MLX-4bit, ezihlanganisa i-GPU, i-CPU, i-Apple Silicon namadivayisi asemaphethelweni.
Ukucubungula umongo omude: Isekela umongo ongu-131,072 tokens, futhi ingacubungula imibhalo emide kanye nemikhondo yama-Agent emijikelezweni eminingi.
Amakhono ezilimi eziningi: Ihlanganisa izilimi ezingu-16, okuhlanganisa isiNgisi, isiShayina, isiJapane nesiKorea, kanye nesi-Arabhu; isichazamazwi samathokheni esingu-128K senzelwe kangcono ukuhlukaniswa kwamathokheni ezilimi ezingezona ezesiLatin.
Ukucabanga okusheshayo: Ku-Apple M5 Max ifinyelela ku-184 tok/s, ishesha kufika izikhathi ezingu-1.7 kune-MoE esesigabeni sayo, futhi idlula ku-21K tok/s ku-B300 yekhadi elilodwa.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokuxhumana lamaphrojekthi e-AI anomthombo ovulekile
Isakhiwo se-MoE sokuvula okuyingxenye: Imodeli inamapharamitha angu-24B isiyonke, kodwa ibhulokhi ngayinye ye-MoE iqukethe ochwepheshe abangu-64. Ngesikhathi sokucabanga, i-top-4 routing ivula ochwepheshe abangu-4 kuphela, okwenza inani langempela lokubala ngethokheni libe ngamapharamitha angu-2.45B kuphela. Lokhu kuyisisekelo samandla amakhulu nezindleko eziphansi zokubala: imodeli igcina lonke ulwazi, kodwa ekuqhubekeni ngakunye isebenzisa ingxenye encane kuphela, ngaleyo ndlela yehlise kakhulu izindleko zokucabanga uma iqhathaniswa namamodeli anenani elifanayo lamapharamitha.
Idizayini yezendlalelo ezixubile: Inethiwekhi enezendlalelo ezingu-40 ye-convolution + GQA yakhiwe ngamabhulokhi angu-30 e-“double-gated short convolution” kanye nezendlalelo ezingu-10 ezixubile ze-GQA. I-short convolution ikhipha ngempumelelo izici zendawo zokulandelana, kuyilapho izendlalelo zokunaka zakha ubudlelwano besikhathi eside. Ukushintshana kwazo kugcina ukubala kwezendlalelo kuhamba kahle futhi kunikeza imodeli amandla omongo omude; lokhu futhi kuyisisekelo sesakhiwo esenza i-LFM isheshe futhi isebenzise inkumbulo encane kumadivayisi asemaphethelweni.
Ukwandiswa kwesichazamazwi samathokheni nokuqeqeshwa komongo omude: Igcina isakhiwo esiyinhloko se-LFM2-24B-A2B yesizukulwane esedlule, kodwa yandisa i-tokenizer isendaweni, yaphinda kabili isichazamazwi sisuka ku-65K saya ku-128K, yathuthukisa kakhulu ukusebenza kwamathokheni ezilimi ezingezona ezesiLatin njengesiShayina nesiJapane. Iwindi lomongo landiswe lisuka ku-32K laya ku-131K; kuhlanganiswe nokuqeqeshwa okumaphakathi okukhethekile komongo omude okusebenzisa ama-tokens angu-300B, lokhu kunikeza imodeli amandla azinzile okucubungula ukulandelana okude.
Inqubo yokuqeqesha enezigaba ezintathu: Ukuqeqeshwa kwangaphambili kusebenzisa cishe ama-tokens angu-33T; ukuqeqeshwa okumaphakathi kuqala kuqinise ukucabanga, izibalo, ikhodi nokulandela imiyalelo kusetshenziswa idatha yekhwalithi ephezulu engama-tokens angu-2T kumongo ongu-32K, bese kugxila emibhalweni emide kusetshenziswa ama-tokens angu-300B kumongo ongu-131K. Zombili lezi zigaba zifaka idatha yokubiza amathuluzi nokusebenzisana kwemijikelezo eminingi. Ukuqeqeshwa kwangemuva kusebenzisa umzila wezigaba ezinhlanu ohlanganisa i-SFT → ukukhethekile kukathisha → ukudluliselwa kolwazi kwamasu amaningi kwe-non-Agent ne-Agent ngemijikelezo emibili → ukuqinisa ukufunda ngaphakathi kwezinhlaka zangempela zama-Agent ezifana ne-OpenClaw ne-Hermes, ukuze imodeli ifunde ngempela “ukucabanga-ukubiza-ukuqinisekisa” endaweni yangempela yamathuluzi.

Thola imodeli: Landa i-LiquidAI/LFM2.5-24B-A2B ku-Hugging Face, bese ukhetha inguqulo yomdabu ye-BF16, FP8, GGUF, ONNX noma i-MLX-4bit ngokuya ngehadiwe.
Faka indawo yokusebenza: Qinisekisa ukuthi indawo ye-Python ine-transformers>=5.0.0, noma usebenzise izinhlaka zokucabanga ezifana ne-vLLM, SGLang, llama.cpp noma i-LM Studio.
Layisha imodeli: Sebenzisa i-AutoModelForCausalLM.from_pretrained() ukuze ulayishe imodeli ngokunemba kwe-bfloat16.
Yakha iziqondiso: Sebenzisa i-tokenizer.apply_chat_template() ukuze kusetshenziswe ngokuzenzakalelayo ifomethi ye-ChatML ezingxoxweni.
Izilungiselelo zamapharamitha: Ngesikhathi sokukhiqiza sebenzisa amapharamitha anconywe ngokusemthethweni athi temperature=0.1, top_k=50, repetition_penalty=1.05; emisebenzini yochungechunge lokucabanga, i-max_new_tokens ingasethwa ibe ngu-4096 noma ngaphezulu.
Hlela ukubizwa kwamathuluzi: Faka uhlu lwamathuluzi ngefomethi ye-JSON ku-system prompt; imodeli izosebenzisa umaka othi <|tool_call_start|> ukukhipha ukubizwa komsebenzi, bese ibuyisela umphumela ngendima ye-tool ngemva kokuwenza.
Xhuma kuhlaka lwe-Agent: Qalisa isevisi endaweni usebenzisa isixhumi esibonakalayo esihambisana ne-OpenAI, bese uyixhuma ku-OpenClaw, Hermes Agent, Pi, OpenCode noma ezinye izinhlaka zama-Agent.

Ukusebenza okuphezulu kakhulu: Emhlabeni wamapharamitha angu-24B, kuvula angu-2.45B kuphela; ku-Apple M5 Max ifinyelela ku-184 tok/s, ishesha kufika izikhathi ezingu-1.7 kune-MoE esesigabeni sayo futhi isebenzisa inkumbulo encane kakhulu.
Okuncane kunqoba okukhulu: Ezivivinyweni zama-Agent namathuluzi ilingana ne-Qwen3.6-35B-A3B enamapharamitha amaningi ngo-50%; i-LiveClaw ithola u-48.09 uma iqhathaniswa no-38.19, kanti ku-AutomationBench idlula indawo yesibili izikhathi ezingu-3.
Ukucabanga kusukela ekuqaleni kuze kube sekugcineni: Imodeli yochungechunge lokucabanga kuphela icabanga kuqala bese iphendula, futhi izindleko eziphansi zokuvula zenza ithokheni ngayinye yokucabanga ibe eshibhile.
Amakhono emvelo e-Agent: Kusukela ekukhishweni kwayo, iqeqeshwe nge-Agentic RL ngaphakathi kwezinhlaka ezidumile ezifana ne-OpenClaw, Hermes, Pi ne-OpenCode; ukubiza amathuluzi kusebenza ngokushesha ngaphandle kokusetha okuningi.
Ivulekile ngaphandle kwemikhawulo: Imodeli eyisisekelo kanye nemodeli yokuqeqeshwa kwangemuva zombili zinezisindo ezivulekile, futhi zingalandwa, zilungiswe noma zisetshenziselwe ukuthengisa ngokukhululeka.
Ukusekelwa kwazo zonke izinkundla: Isekela ngokwendabuko i-llama.cpp, MLX, vLLM, SGLang ne-ONNX, futhi iyahambisana nehadiwe ye-Apple, AMD, Qualcomm ne-NVIDIA.

Ilabhulali yamamodeli ye-HuggingFace: https://huggingface.co/LiquidAI/LFM2.5-24B-A2B

I-Agent yangasese esebenza kudivayisi: Qalisa i-Agent yasendaweni ubusuku nemini kumadivayisi omuntu afana ne-MacBook ukuze yenze ukuhlela, ukubiza amathuluzi nokwenza imisebenzi; idatha ayishiyi idivayisi.
Umsizi ongaxhunyiwe ku-inthanethi/osemaphethelweni: Sebenzisa amafomethi e-GGUF, ONNX noma e-MLX ezindaweni lapho inethiwekhi ingazinzile noma ingekho, njengasezimotweni, endle noma kumatheminali ashumekiwe, ukuze unikeze ukucabanga okungaxhunyiwe ku-inthanethi.
Ukuphathwa kwamathuluzi okuhlakaniphile: Yisebenzise njengemodeli eshayela izinhlaka ezifana ne-OpenClaw, Hermes Agent ne-OpenCode ukuze yenze ukugeleza kwemisebenzi enezinyathelo eziningi, njengokukhiqiza imibiko ehleliwe, ukuhlela amafayela nokusebenza ngekhodi.
Ukuzenzakalela kwezinqubo zebhizinisi: Sebenzisa amandla ayo ku-ToolSandbox naku-τ³-Bench Banking ukuze iphathe izingxoxo zebhange, zokuqasha nezama-oda ezidinga ukubuza amasistimu nokwenza ukuqinisekiswa kwemijikelezo eminingi.