Ndesịta ozi ndị ahụ
Báà
Lance 3B (unified) bụ a Model AI e mepụtara site na ByteDance. Ọrụ na Cross-task research, prototyping pipelines that need image + video + edit + VQA from one model, "one model, four tasks" demos. Apache 2.0, commercial use OK.. Self-hosted na Free.ai GPUs - na-agba ọsọ n'efu megide ụbọchị gị token pool (100 tokens Oge ọrụ). E wepụtara ya n'okpuru Apache 2.0 — iji azụmahịa ekwenyela na Free.ai.
Jiri site na API
curl https://api.free.ai/v1/chat/ \
-H "Authorization: Bearer YOUR_KEY" \
-d '{"model":"lance-3b"}'
Ajụjụ ndị na-emekarị
Lance bụ ByteDance 2025 native unified multimodal model - 3B active parameters n'okpuru Apache 2.0. Otu set nke weights na-ekpuchi ọrụ anọ: ngwe→ịhụnanya (768×768), ihumkpụrụedemede (768×768), ngwe→video (480p, ruo 121 frames ≈ 5 sekọnd), na ihumkpụrụedemede + video nghọta (VQA, captioning). E mepụtara na Qwen-derived LLM backbone na Wan-Video VAE na Qwen.5-VL ViT. Self-hosted na Free.ai's H200 na enweghị upstream provider, enweghị API markup, na enweghị ụgwọ ọ bụla n'ime gị token balance.
Ọtụtụ ndị na-emepe stacks na-ahọrọ onye ọkachamara kacha mma maka elu nke ọ bụla - SDXL ma ọ bụ FLUX maka ịmepụta inyogo raw, Qwen-Image-Edit maka ntọala, Wan 2.2 maka vidiyo, Qwen.5-VL maka nghọta-asụsụ. Lance na-eweta ihe nketa nke ọrụ-ọdịnaya maka ntọala ọrụ-ọdịnaya: ihe ngosi nkeonwe na-enye ihe niile, yabụ na inyogo ị na-emepụta ma ọ bụ gbanwee na-echekwa ụdị ya, na VQA nke model na-enye banyere video na-abịarute asụsụ model na ebe nnyocha ahụ. Ọ bara uru maka nyocha na ndepụta nke na-eweta uru site na otu model dị n'otu kama pipeline nke anọ.
Choose Lance mgbe: ịchọrọ ụdị dị iche iche n'ime inyogo + edit + vidiyo site na otu ụdị, ị na-enyocha multi-task pipeline na "otu ụdị" angle ihe ndị dị mkpa, ma ọ bụ ịchọrọ ikike ikike na unified workflow. Choose specialists mgbe: ịchọrọ ogo dị elu nke raw image gen (FLUX.2 Klein > Lance na> 768²), ogologo / elu-ọdịnaya vidiyo (Wan 2.2 TI2V-5B ma ọ bụ HunyuanVideo > Lance na> 480p), ma ọ bụ VQA dị ngwa ngwa na nkata (Qwen.5-VL bụ mgbe niile na-ekpo ọkụ na H200, Lance nwere ike ịmalitegharị).
Text→image na image-edit: 5,000 tokens (na-ahazi FLUX-class image gen). Text→video: 15,000 tokens (na-ahazi CogVideoX / Wan 5B class). Image+video VQA: 1,000 tokens. The higher cost vs SDXL (1,000) reflects Lance's heavier cold-load — every call evicts the remainder of the warm fleet and re-loads 40 GB of weights, which adds 25-40 s on top of the inference itself. We're billing for total wall-clock GPU time, not just inference.
Mgbe cold-load (~ 25-40 s): image gen ~ 12-20 s, image edit ~ 15-25 s, text→video ~ 60-180 s (dabere na num_frames), VQA ~ 3-8 s. Ọtụtụ Lance na-akpọ cold-loads model n'ihi na ọ gaghị ekwe omume na-ejikọta ya na ndị ọzọ na-ekpo ọkụ na H200, yabụ na-ekpo ọkụ-ọgwụgwụ bụ akụkụ nke ekwentị niile, ọ bụghị naanị nke mbụ.
Ihu nkesa na nhazi inyogo a kpụgharịrị na 768×768. Ihu nkesa a kpụgharịrị na 480p (dị ka 480×848 landscape) na kapped na 121 frames (~5 sekọnd na 24 fps). Ha bụ ihenhọrọ Lance kpụgharịrị na; ịkpụzi elu chọrọ upscaling site na móòdù dị iche iche (nwale /image/upscaler/ maka inyogo mọọbụ /video/upscaler/ maka vidio).
Janus (DeepSeek) na Gosi-o kpụga nghọta na mmepe n'ime isi dị iche iche na backbone nke e mepụtara; Lance bụ nke dị n'otu n'ụzọ dị mfe - otu setịpụrụ nke mmepe+nghọta n'ime isi na ọrụ token ndị dị n'ime. Emu3 (BAAI) na-egosipụta ihe niile dị ka ihe dị iche iche gụnyere pikseli, nke na-enye ya mmepe na-atọ ọchị ma ọ bụ nke dị ala na nrụnye. Lance's pitch bụ ihe n'ime ọrụ anọ na 3B params nakwa ya Wan-derived VAE nke na-elekọta vidiyo na-enweghị atụ (Janus na Gosi-o bụ inyogo-ọbụla).
Apache 2.0 — ọbụna na-arọ (huggingface.co/bytedance-research/Lance) na GitHub repo (github.com/bytedance/Lance). Enweghị nkwụsị nke mpaghara, enweghị MAU cap, enweghị onye na-enweghị ọrụ, enweghị nchọpụta-ọbụla clause. Ọrụ bụ gị iji rụọ ọrụ na-enweghị ikike ma ọ bụ ihe ndị chọrọ n'ebe ọ bụla n'ebe ahụ Apache 2.0 ngwe ikike.
40 GB minimum kwa ByteDance's README. 3B params na-arụ ọrụ bụ ndị na-emebi emebi - Qwen LLM + Wan VAE + Qwen.5-VL ViT niile na-abanye n'ime nghọta. Iji rụpụta onwe gị ị ga-achọ A100 80 GB, A6000 48 GB, ma ọ bụ H100 / H200 na 40 GB ọbụlagodi. Anyị na-agba ya na H200 anyị (141 GB zuru oke) mana ọ na-ewepụkwa ndị ọzọ nke ụdị ndị a na-ebugharị n'otu oku n'ihi na ọ bụ nnukwu ọdụ na igbe ahụ.
Ya - POST JSON mọọbụ multipart ka /v1/multimodal/lance/ na api.free.ai na {task: "t2i" | "image_edit" | "t2v" | "vqa", pịịpị: "...", inyogo: <upload> mọọbụ inyogo_url: "/static/outputs/..."}. Ogoozi nke onye na-ebubata site na kii debanye aha API. Nnọọ na-agụnye job_id, ihenhọrọ URL, na share_token. /api/ nwere ihenhọrọ curl kwa ọrụ.
Anyị na-egosi Lance na-enyocha n'ihi na ọ bụghị ihe dị mma maka okporo ụzọ dị elu - ọ bụla na-akpọ na-ewepụ ihe ọkụkụ na-ekpo ọkụ na-ekpo ọkụ. Anyị nwere ike ịgbakwunye "ọnụ ọkụ Lance" n'oge na-adịghị anya ma ọ bụrụ na iji na-egosipụta ịrụ ọrụ na-egosipụta slot, ma ọ bụ anyị nwere ike ịgbakwunye nke abụọ H200 n'ụzọ pụrụ iche maka ụdị ndị a na-ejikọta. Maka ugbu a ọ dị na ntinye ego nke Free.ai's self-hosted models na enweghị nkwụghachi ụgwọ, naanị ọnụ ahịa nke ọnụ ahịa nke na-egosipụta oge GPU nke wall-clock.
Igosi ndị a na-ebubata maka nhazi inyogo na VQA a na-ehichapụ n'oge na-adịghị anya mgbe ọrụ ahụ gasịrị. A na-ebubata ihenhọrọ ndị ahụ na CDN anyị maka awa 24 (ụbọchị 7 maka ndị ọrụ na-akwụ ụgwọ) ka ị wee nwee ike ibudata ya na /account/?tab=history. E nweghị ihe a na-ekerịta na ByteDance — ihenhọrọ ndị ahụ na-arụ ọrụ na mpaghara ebe anyị na haadị. Ndepụta zuru ezu na /privacy/.