{"version":1,"type":"story","url":"https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization","json":"https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization.json","markdown":"https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization.md","slug":"llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization","headline":"LLaMa Adds Maple 20B-A1B MoE Architecture with Ternary Quantization","summary":"AlexGabbia, a contributor to the LLaMa project, has added the Maple 20B-A1B ternary Multi-Expert (MoE) architecture to the LLaMa codebase. This update includes support for both q20 and tQ10/TQ20 quantization schemes. The team ran tests on a RTX 5070 Ti GPU, finding that TQ20 performed slightly worse than q20 in perplexity scores across different contexts. They also noted differences between CPU and CUDA backends for the same model file, attributing these to backend-specific optimizations rather than quantization or architecture issues. The team is now working on aligning this release with other updates and preparing documentation for the new Maple 20B-A1B MoE architecture.","keyPoints":["LLaMa adds support for Maple 20B-A1B ternary MoE architecture","TQ10/TQ20 quantization scheme included in the update","CPU and CUDA backends show different performance metrics"],"whyItMatters":"This update is significant as it expands LLaMa's capabilities to support a new type of model architecture, which could be useful for improving efficiency or flexibility in AI applications.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":[],"models":["LLaMa","Maple"],"people":["AlexGabbia"]},"firstPublishedAt":"2026-09-14T12:11:18Z","updatedAt":"2026-09-14T12:11:18Z","sourceCount":1,"hasPrimarySource":true,"sources":[{"outlet":"github.com","title":"llama: add Maple 20B-A1B ternary MoE architecture (CPU) by AlexGabbia · Pull Request #27000 · ggml-org/llama.cpp","url":"https://github.com/ggml-org/llama.cpp/pull/27000","publishedAt":"2026-09-14T12:11:18Z","type":"primary","primary":true,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1wg1o5b/llama_add_maple_20ba1b_ternary_moe_architecture/","points":null}],"thread":null,"cite":{"text":"Digest AI, \"LLaMa Adds Maple 20B-A1B MoE Architecture with Ternary Quantization\", 14 September 2026, https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization","publisher":"Digest AI","title":"LLaMa Adds Maple 20B-A1B MoE Architecture with Ternary Quantization","datePublished":"2026-09-14T12:11:18Z","url":"https://digestai.news/story/llama-adds-maple-20b-a1b-moe-architecture-with-ternary-quantization"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}