{"id":182824,"date":"2026-08-11T04:54:15","date_gmt":"2026-08-11T02:54:15","guid":{"rendered":"https:\/\/prohoster.info\/blog\/novosti-interneta\/yazykovaya-model-s-289-mln-parametrov-zarabotala-na-esp32-s3"},"modified":"2026-08-12T18:21:41","modified_gmt":"2026-08-12T16:21:41","slug":"yazykovaya-model-s-289-mln-parametrov-zarabotala-na-esp32-s3","status":"publish","type":"post","link":"https:\/\/prohoster.info\/pt\/blog\/news\/yazykovaya-model-s-289-mln-parametrov-zarabotala-na-esp32-s3","title":{"rendered":"Modelo de linguagem com 28,9 milh\u00f5es de par\u00e2metros foi implementado no ESP32-S3","gt_translate_keys":[{"key":"rendered","format":"text"}]},"content":{"rendered":"

<\/i><\/a>
\nO desenvolvedor Vyacheslav Serbov, que atua sob o pseud\u00f4nimo slvDev, demonstrou<\/a> desenvolveu uma gera\u00e7\u00e3o de texto completamente local no microcontrolador ESP32-S3. O modelo de linguagem criado por ele cont\u00e9m 28,9 milh\u00f5es de par\u00e2metros e gera cerca de 9,9 tokens por segundo, sem acessar um servidor ou outros recursos computacionais externos. As hist\u00f3rias curtas geradas s\u00e3o exibidas em um OLED conectado.<\/p>\n

 <\/p>\n

Projeto esp32-ai<\/a> testado no m\u00f3dulo ESP32-S3 N16R8, que possui 512 KB de SRAM embutida, 8 MB de PSRAM e 16 MB de mem\u00f3ria flash. O modelo quantizado para quatro bits ocupa 14,9 MB. O reposit\u00f3rio cont\u00e9m o c\u00f3digo-fonte do ambiente de execu\u00e7\u00e3o em C, ferramentas de treinamento e quantiza\u00e7\u00e3o do modelo, firmware, testes e scripts para carreg\u00e1-lo na placa.<\/p>\n

 <\/p>\n

Os 28,9 milh\u00f5es de par\u00e2metros reivindicados n\u00e3o devem ser comparados diretamente ao n\u00famero de par\u00e2metros dos LLMs de desktop convencionais. De acordo com o relat\u00f3rio detalhado do desenvolvedor<\/a>, o modelo consiste em um n\u00facleo computacional denso de aproximadamente 559 mil par\u00e2metros, uma camada de sa\u00edda de 3,1 milh\u00f5es de par\u00e2metros e uma tabela de representa\u00e7\u00f5es vetoriais em camadas de cerca de 25 milh\u00f5es de par\u00e2metros. \u00c9 esta \u00faltima que fornece a maior parte do tamanho formal do modelo, mas n\u00e3o \u00e9 processada inteiramente ao gerar cada token.<\/p>\n

 <\/p>\n

Para hospedagem do modelo, tr\u00eas n\u00edveis de mem\u00f3ria s\u00e3o utilizados. Os dados mais frequentemente utilizados est\u00e3o na SRAM interna, a camada de sa\u00edda, o cache KV e os buffers tempor\u00e1rios est\u00e3o na PSRAM, enquanto a tabela de 25 milh\u00f5es permanece na mem\u00f3ria flash e est\u00e1 acess\u00edvel via mapeamento no espa\u00e7o de endere\u00e7o. Para cada token, apenas cerca de seis linhas s\u00e3o lidas, totalizando aproximadamente 450 bytes.<\/p>\n

 <\/p>\n

Esse dispositivo \u00e9 baseado na tecnologia Per-Layer Embeddings, aplicada pelo Google no Gemma 3n. De acordo com a documenta\u00e7\u00e3o do Google<\/a>, os par\u00e2metros PLE podem ser armazenados fora da mem\u00f3ria de trabalho do modelo e adicionados ao processo de sa\u00edda conforme os diferentes camadas s\u00e3o processadas. No esp32-ai, essa abordagem foi transferida para a hierarquia de mem\u00f3ria do microcontrolador, onde a SRAM r\u00e1pida \u00e9 particularmente limitada.<\/p>\n

 <\/p>\n

A primeira vers\u00e3o funcional do ambiente de execu\u00e7\u00e3o escrito em C fornecia apenas 0,57 tokens por segundo. Ap\u00f3s a coloca\u00e7\u00e3o da camada de sa\u00edda na PSRAM, a transi\u00e7\u00e3o para ativa\u00e7\u00f5es de oito bits, a distribui\u00e7\u00e3o de c\u00e1lculos entre os dois n\u00facleos Xtensa LX7 e outras otimiza\u00e7\u00f5es, a lat\u00eancia foi reduzida para 94,9 ms por etapa do modelo. A velocidade medida final alcan\u00e7ou 9,88 tokens por segundo, levando em conta todo o processo de gera\u00e7\u00e3o. O principal limite agora \u00e9 a largura de banda da PSRAM ao ler a camada de sa\u00edda.<\/p>\n

 <\/p>\n

O modelo foi treinado em um conjunto de dados sint\u00e9tico TinyStories<\/a>, consistindo em hist\u00f3rias simples em ingl\u00eas com vocabul\u00e1rio limitado. Portanto, ele \u00e9 capaz de continuar frases e compor pequenas hist\u00f3rias coerentes, mas n\u00e3o \u00e9 destinado a responder perguntas, executar instru\u00e7\u00f5es, programar ou reproduzir conhecimentos factuais. O desenvolvedor v\u00ea o trabalho principalmente como uma demonstra\u00e7\u00e3o da coloca\u00e7\u00e3o eficaz do modelo na mem\u00f3ria de um microcontrolador, e n\u00e3o como um chatbot aut\u00f4nomo pronto para uso.<\/p>\n

 <\/p>\n

Na documenta\u00e7\u00e3o atual, o autor enfatiza que o esp32-ai \u00e9 um desenvolvimento aut\u00f4nomo. Os projetos llama2.c<\/a> de Andrei Karpaty e o lan\u00e7amento anterior do modelo de 260 mil par\u00e2metros no ESP32-S3 s\u00e3o citados apenas como refer\u00eancias e exemplos de trabalhos anteriores: o c\u00f3digo, os pontos de verifica\u00e7\u00e3o e a metodologia n\u00e3o foram diretamente extra\u00eddos deles. O c\u00f3digo-fonte do esp32-ai \u00e9 distribu\u00eddo sob a licen\u00e7a MIT<\/a>.<\/p>\n

 <\/p>\n

Fonte: linux.org.ru<\/a><\/p>","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"excerpt":{"rendered":"

\u0420\u0430\u0437\u0440\u0430\u0431\u043e\u0442\u0447\u0438\u043a \u0412\u044f\u0447\u0435\u0441\u043b\u0430\u0432 \u0421\u0435\u0440\u0431\u043e\u0432, \u0432\u044b\u0441\u0442\u0443\u043f\u0430\u044e\u0449\u0438\u0439 \u043f\u043e\u0434 \u043f\u0441\u0435\u0432\u0434\u043e\u043d\u0438\u043c\u043e\u043c slvDev, \u043f\u0440\u043e\u0434\u0435\u043c\u043e\u043d\u0441\u0442\u0440\u0438\u0440\u043e\u0432\u0430\u043b \u043f\u043e\u043b\u043d\u043e\u0441\u0442\u044c\u044e \u043b\u043e\u043a\u0430\u043b\u044c\u043d\u0443\u044e \u0433\u0435\u043d\u0435\u0440\u0430\u0446\u0438\u044e \u0442\u0435\u043a\u0441\u0442\u0430 \u043d\u0430 \u043c\u0438\u043a\u0440\u043e\u043a\u043e\u043d\u0442\u0440\u043e\u043b\u043b\u0435\u0440\u0435 ESP32-S3. \u0421\u043e\u0437\u0434\u0430\u043d\u043d\u0430\u044f \u0438\u043c \u044f\u0437\u044b\u043a\u043e\u0432\u0430\u044f \u043c\u043e\u0434\u0435\u043b\u044c \u0441\u043e\u0434\u0435\u0440\u0436\u0438\u0442 28,9 \u043c\u043b\u043d \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 \u0438 \u0432\u044b\u0434\u0430\u0451\u0442 \u043e\u043a\u043e\u043b\u043e 9,9 \u0442\u043e\u043a\u0435\u043d\u0430 \u0432 \u0441\u0435\u043a\u0443\u043d\u0434\u0443, \u043d\u0435 \u043e\u0431\u0440\u0430\u0449\u0430\u044f\u0441\u044c \u043a \u0441\u0435\u0440\u0432\u0435\u0440\u0443 \u0438\u043b\u0438 \u0434\u0440\u0443\u0433\u0438\u043c \u0432\u043d\u0435\u0448\u043d\u0438\u043c \u0432\u044b\u0447\u0438\u0441\u043b\u0438\u0442\u0435\u043b\u044c\u043d\u044b\u043c \u0440\u0435\u0441\u0443\u0440\u0441\u0430\u043c. \u0421\u0433\u0435\u043d\u0435\u0440\u0438\u0440\u043e\u0432\u0430\u043d\u043d\u044b\u0435 \u043a\u043e\u0440\u043e\u0442\u043a\u0438\u0435 \u0440\u0430\u0441\u0441\u043a\u0430\u0437\u044b \u0432\u044b\u0432\u043e\u0434\u044f\u0442\u0441\u044f \u043d\u0430 \u043f\u043e\u0434\u043a\u043b\u044e\u0447\u0451\u043d\u043d\u044b\u0439 OLED-\u0434\u0438\u0441\u043f\u043b\u0435\u0439.   \u041f\u0440\u043e\u0435\u043a\u0442 esp32-ai \u0438\u0441\u043f\u044b\u0442\u0430\u043d \u043d\u0430 \u043c\u043e\u0434\u0443\u043b\u0435 ESP32-S3 N16R8, \u0440\u0430\u0441\u043f\u043e\u043b\u0430\u0433\u0430\u044e\u0449\u0435\u043c […]<\/p>\n","protected":false,"gt_translate_keys":[{"key":"rendered","format":"html"}]},"author":9,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[702],"tags":[],"class_list":["post-182824","post","type-post","status-publish","format-standard","hentry","category-news"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\n","aioseo_head_json":{"title":"\ud83e\udd47O modelo de linguagem com 28,9 milh\u00f5es de par\u00e2metros funcionou no ESP32-S3 | ProHoster","description":".","canonical_url":"https:\/\/prohoster.info\/pt\/blog\/news\/yazykovaya-model-s-289-mln-parametrov-zarabotala-na-esp32-s3","robots":"max-image-preview:large","keywords":"","webmasterTools":{"miscellaneous":""},"schema":null,"og:locale":"pt_PT","og:site_name":"ProHoster | \u041a\u0443\u043f\u0438\u0442\u044c \u043d\u0430\u0434\u0435\u0436\u043d\u044b\u0439 \u0445\u043e\u0441\u0442\u0438\u043d\u0433 \u0434\u043b\u044f \u0441\u0430\u0439\u0442\u043e\u0432 \u0441 \u0437\u0430\u0449\u0438\u0442\u043e\u0439 \u043e\u0442 DDoS, VPS VDS \u0441\u0435\u0440\u0432\u0435\u0440\u044b","og:type":"article","og:title":"\ud83e\udd47\u042f\u0437\u044b\u043a\u043e\u0432\u0430\u044f \u043c\u043e\u0434\u0435\u043b\u044c \u0441 28,9 \u043c\u043b\u043d \u043f\u0430\u0440\u0430\u043c\u0435\u0442\u0440\u043e\u0432 \u0437\u0430\u0440\u0430\u0431\u043e\u0442\u0430\u043b\u0430 \u043d\u0430 ESP32-S3 | ProHoster","og:description":".","og:url":"https:\/\/prohoster.info\/pt\/blog\/news\/yazykovaya-model-s-289-mln-parametrov-zarabotala-na-esp32-s3","og:image":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:secure_url":"https:\/\/prohoster.info\/wp-content\/uploads\/2021\/11\/logo-350.jpg","og:image:width":350,"og:image:height":350,"article:published_time":"2026-08-11T02:54:15+00:00","article:modified_time":"2026-08-12T16:21:41+00:00","article:publisher":"https:\/\/www.facebook.com\/prohoster"},"aioseo_meta_data":{"post_id":"182824","title":null,"description":null,"keywords":null,"keyphrases":{"focus":[],"additional":[]},"primary_term":null,"canonical_url":null,"og_title":null,"og_description":null,"og_object_type":"default","og_image_type":"default","og_image_url":null,"og_image_width":null,"og_image_height":null,"og_image_custom_url":null,"og_image_custom_fields":null,"og_video":null,"og_custom_url":null,"og_article_section":null,"og_article_tags":null,"twitter_use_og":false,"twitter_card":"default","twitter_image_type":"default","twitter_image_url":null,"twitter_image_custom_url":null,"twitter_image_custom_fields":null,"twitter_title":null,"twitter_description":null,"schema":{"blockGraphs":[],"customGraphs":[],"default":{"data":{"Article":[],"Course":[],"Dataset":[],"FAQPage":[],"Movie":[],"Person":[],"Product":[],"ProductReview":[],"Car":[],"Recipe":[],"Service":[],"SoftwareApplication":[],"WebPage":[]},"graphName":"","isEnabled":true},"graphs":[]},"schema_type":"default","schema_type_options":null,"pillar_content":false,"robots_default":true,"robots_noindex":false,"robots_noarchive":false,"robots_nosnippet":false,"robots_nofollow":false,"robots_noimageindex":false,"robots_noodp":false,"robots_notranslate":false,"robots_max_snippet":null,"robots_max_videopreview":null,"robots_max_imagepreview":"large","priority":null,"frequency":null,"local_seo":null,"seo_analyzer_scan_date":null,"breadcrumb_settings":null,"limit_modified_date":false,"reviewed_by":null,"ai":null,"created":"2026-08-11 06:43:49","updated":"2026-08-11 06:43:49","focus_keyword":null,"additional_keywords":null,"truseo_locale":null},"gt_translate_keys":[{"key":"link","format":"url"}],"_links":{"self":[{"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/posts\/182824","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/users\/9"}],"replies":[{"embeddable":true,"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/comments?post=182824"}],"version-history":[{"count":1,"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/posts\/182824\/revisions"}],"predecessor-version":[{"id":182897,"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/posts\/182824\/revisions\/182897"}],"wp:attachment":[{"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/media?parent=182824"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/categories?post=182824"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/prohoster.info\/pt\/wp-json\/wp\/v2\/tags?post=182824"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}