{
"$type": "site.standard.document",
"bskyPostRef": {
"cid": "bafyreieeyoir26sjr4wiq56c63er5wnhydb3hg4rlbfftv6exqp27k4hca",
"uri": "at://did:plc:nfto3lv2rcs5s7h7digotzlu/app.bsky.feed.post/3mojy4twa77t2"
},
"coverImage": {
"$type": "blob",
"ref": {
"$link": "bafkreig7saeenw6c3prrms24e2bpcmq7mskq26az4ajg73ufx33cf4gf2a"
},
"mimeType": "image/png",
"size": 32236
},
"path": "/packages/juice_llm_llamacpp",
"publishedAt": "2026-06-18T04:22:19.025Z",
"site": "https://pub.dev",
"textContent": "Embedded on-device LLM runtime for juice_llm — an LlamaCppProvider backed by llama.cpp (GGUF, Metal) via llama_cpp_dart. Changelog excerpt: ### Added - **`chatFormat`option + `gemmaChatFormat`**— for models whose embedded chat template llama.cpp can't apply (Gemma 4 ships tool-use Jinja the runtime can't parse). When set, the prompt is built manually and generation uses the raw session path instead of `chat.generate`. `gemmaChatFormat`handles the `user … model`format (no system role — system folds into the first user turn). Verified end-to-end with Gemma 4 E2B; the default (embedded-templa[...]",
"title": "v0.1.1 of juice_llm_llamacpp",
"updatedAt": "2026-06-18T04:20:28.630Z"
}