llama.cpp Hits 1.0 and Promises No More Surprise Breakage. Getting There Broke 36 Functions, and the Release Closes Server Holes Older Builds Still Have.
ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on November 13, the project's first semantic-versioned release: from now on a major bump means a breaking change to the public C API. The jump itself is a break, removing llama_set_warmup and 35 deprecated functions, the defrag threshold flag, and clip.h. The server notes are a list of security fixes, for CORS origins, DNS rebinding, CORS proxy and media URL SSRF, child instance access, and an HF_TOKEN now restricted to huggingface.co, plus new opt-ins for unauthenticated tools on network addresses and private media URLs. The release also adds an FP8 KV cache, a host-memory MoE expert cache, Kimi K3.1 Pro, and a speech endpoint. No CVE identifiers were listed. We read the release notes and the blog.
By T.W. Ghost
ggml-org published llama.cpp v1.0.0 today, November 13, 2026, with ggml v1.0.0 alongside it, the first release of either under semantic versioning after three and a half years and more than 150 architecture variants. The promise is simple: a major version bump means a breaking change to the public C API in ./include, a minor bump means features and models, a patch means fixes. Anyone who has built on llama.cpp knows what that promise is worth, because until today a routine pull could change a function signature.
The first break is the release itself. Version 1.0.0 removes llama_set_warmup and 35 deprecated functions, drops defrag_thold and the --defrag-thold flag, folds clip.h into mtmd, removes the non-fused GDN paths, and promotes the NextN API into llama.h. And the server section reads like an advisory. It fixes handling of --cors-origins, DNS rebinding, SSRF through the CORS proxy and media URLs, and child instance access; restricts HF_TOKEN to huggingface.co hosts; and adds two opt-in flags, --allow-unauthenticated-tools and --media-allow-private-urls, for behavior that used to be on by default on network addresses. No CVE identifiers appear in what we read. Anyone running llama-server on a reachable address should read this as a security upgrade first and a version milestone second. We read both release notes and the Hugging Face blog post. Here is what held up.
What Changed
| llama.cpp v1.0.0 and ggml v1.0.0 (November 13, 2026) | |
|---|---|
| Published | llama.cpp 10:20 UTC; ggml 08:34 UTC |
| Versioning | MAJOR = breaking public C API change in ./include; MINOR = features and models; PATCH = fixes |
| Removed | llama_set_warmup; 35 deprecated functions; defrag_thold and --defrag-thold; clip.h (into mtmd); non-fused GDN paths |
| Promoted | NextN API into llama.h |
| Added | OCP FP8 KV cache; host-memory MoE expert cache with CPU share; external PLE shards |
| New models | Kimi K3.1 Pro (K3 with full-attention MLA replaced by DSA); Ripple; xing4_0; Kimi K3.1 vision; drama-3 TTS; /v1/audio/speech |
| Server security | Fixes for --cors-origins, DNS rebinding, CORS proxy SSRF, media URL SSRF, child instance access; HF_TOKEN restricted to huggingface.co; opt-ins --allow-unauthenticated-tools and --media-allow-private-urls |
| CVEs | None listed in what we read |
1.0 is a breaking upgrade. For direct users of the C API, the first semver release is the one that removes 36 functions. Read the list before you bump.
Older builds keep the holes. The server fixes ship in 1.0; a build from last month has the behavior they fix. Without CVE identifiers, scanners will not flag it for you.
Network-exposed tools now need a flag. Setups that relied on unauthenticated tool calls or private media URLs over a network address stop working until the opt-in is set. That is the right default and a silent change for anyone who does not read the notes.
What we did not check. Downstream tools such as Ollama and LM Studio bundle their own builds. Whether and when they carry these fixes is a question for their release notes, not an assumption to make from this one.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| v1.0.0 published Nov 13, 2026, both projects | Verified (GitHub releases) |
| Semver policy | Verified (Hugging Face blog) |
| Removals, promotions, additions, new models | Verified (release notes) |
| Server security fixes and opt-in flags | Verified (release notes) |
| 150+ architectures; 3.5 years | Vendor (blog) |
| CVE identifiers | Not found |
| Downstream exposure (Ollama, LM Studio) | Not checked |
What To Do
Upgrade any network-reachable llama-server now, then review --cors-origins and the two new tool and media flags against what your setup relied on.
Skip the blind bump if you link the library; check the removed-function list first, and expect the next breaking change to come with a major version number instead of a commit message.
The frontier is elsewhere: the model this release adds support for, Kimi K3.1 Pro, is the open-weights leader on Artificial Analysis's board, and Muse Spark Open remains Meta's open claim on the same board. Our model comparison has the current standings side by side.
Sources
*Not sure whether local inference or an API fits your workload? Take the free 2-minute quiz and get matched. Then read about the open model whose weights this engine now loads: Kimi K3.1 Pro.*