Skip to content

Toolsggml-org

llama.cpp and ggml reach 1.0: semantic versioning, a host-memory MoE expert cache, and 35 deprecated functions removed

llama.cpp at a glance

ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on November 13, 2026, giving the local-inference runtime a stable versioning contract for the first time: breaking public C API changes now mean a major version bump, and 1.0 removes 35 deprecated functions while adding a host-memory cache for MoE experts.

Ask your AI about this story

Use this story in your business: llama.cpp and ggml reach 1.0: semantic versioning, a host-memory MoE expert cache, and 35 deprecated functions removed (ggml-org). Ask your AI to explain what happened, how it could affect your business and how to put it to work. It also weighs price, speed, intelligence and the lab behind it. The link to the story is included.

What changed with llama.cpp

ggml-org released llama.cpp v1.0.0 on November 13, 2026, alongside ggml v1.0.0, and explained the milestone in a Hugging Face blog post from the ggml-org organization. Version 1.0 adopts semantic versioning for the public C API: a major bump signals a breaking change in the ./include API, a minor bump adds backward-compatible features, backends, models or APIs, and a patch bump is a bug fix. Until now the project versioned by build number, so downstream projects had no signal for when an upgrade would break their integration.

The 1.0 release removes the deprecated public API, including llama_set_warmup and 35 deprecated functions in total, along with the defrag_thold setting, and promotes the NextN and speculative-decoding API into llama.h. For running large mixture-of-experts models, it adds a host-memory MoE expert cache with a configurable CPU share (moe_cache_size and moe_cache_cpu_share), aimed at models larger than GPU memory, plus OCP FP8 KV caches and external PLE shards. The release also adds support for new models and hardens the built-in server, including SSRF, CORS and DNS-rebinding protections, MCP cancellation, and a /v1/audio/speech endpoint.

llama.cpp in numbers

Values are copied from the linked source on the date shown. They are not IA Compa measurements.

Deprecated public API functions removed in llama.cpp 1.0as of · Source for Deprecated public API functions removed in llama.cpp 1.0: github.com (opens in a new tab)
35

Where to compare llama.cpp

Independent leaderboards and official resources for comparing this model yourself. Rankings change often; each site shows its own current data.

Why llama.cpp matters for businesses

Why it matters: llama.cpp is the runtime underneath a large share of local and self-hosted AI deployments, including tools a small business can run on its own hardware to keep client data in-house. A stable versioning contract makes those deployments maintainable: you can pin a major version, plan upgrades, and know a minor release will not break your integration. The MoE expert cache matters for cost, because it makes very large open models usable on machines whose GPU memory alone could not hold them, which lowers the hardware bill for private, on-premise AI.

Put it to work: AI services

Get AI News by email

Verified AI stories, each with its official source, on Monday and Thursday mornings. Add breaking news to hear about the biggest ones as soon as we publish them.

Every story links its official source

Free, in English or Spanish

Unsubscribe with one click, any time

What to send

Sources for this story

  1. Official
  2. Official

Last verified: . IA Compa summarizes public announcements for business readers. We are not affiliated with the companies mentioned. Always confirm details, pricing, and availability with the official source.

  • #llama-cpp
  • #ggml
  • #local-inference
  • #open-source
  • #semantic-versioning
  • #mixture-of-experts
  • #self-hosted-ai

Want this working in your business?

We turn new AI capabilities into agents, voice receptionists, and automations you can actually use.