Inter-office memo
- To
- Everyone running models on their own computer
- From
- Mo, Chief Executive
- Date
- November 15, 2026, 12:31 PM CST
- Re
- llama.cpp v1.0.0
llama.cpp 1.0 is out. What’s new, what breaks, and why to update llama-server
On November 13, 2026, ggml-org released llama.cpp v1.0.0, and from now on a breaking change to its C API bumps the major version. It runs Kimi K3.1 Pro and Ripple-65B Preview, adds FP8 KV caches and a GPU cache for MoE experts, and tightens llama-server, which now won’t start with tools on a network address and no API key. Scripts that pass --defrag-thold must drop it.
What ggml-org released
On November 13, 2026, Georgi Gerganov, who started llama.cpp, posted: “llama.cpp v1.0.0 is now available.” It came with ggml v1.0.0, the library underneath it. The llama.cpp team at Hugging Face, which maintains the project, said what the number means: “Moving forward, both llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps.”
| Version change | When it happens, in ggml-org’s words |
|---|---|
| Major, such as 2.0.0 | “Breaking change to the public C API in ./include” |
| Minor, such as 1.1.0 | “Backward-compatible feature, backend updates, model support, or API addition” |
| Patch, such as 1.0.1 | “Bug fix with no API change” |
That promise is mostly for the apps built on llama.cpp. The release notes cover everything since v0.9.0, and the numbered nightly builds carry on beside the releases, as ggml-org’s versioning page describes. Version 1.0.0 is the same code as nightly build b12359, which is where the release’s binaries are. Hugging Face also says llama.cpp now supports “more than 150 architecture variants.”
What’s new if you run models
| What | What it does | How to use it |
|---|---|---|
| Kimi K3.1 Pro, Ripple-65B Preview, xing4_0 | New text models | Load their GGUF files |
| Kimi K3.1 and MageVL-Cowork vision, drama-3 speech | New image and voice models | Through llama.cpp’s multimodal support |
/v1/audio/speech |
An OpenAI-compatible speech endpoint in llama-server | Send it text, as you would OpenAI’s |
| FP8 KV cache | Keeps the context cache in 8-bit floats, half the size of the default f16 | -ctk f8_e4m3 -ctv f8_e4m3 |
| MoE expert cache | Keeps a mixture-of-experts model’s busiest experts on the GPU while the rest sit in system memory | -cmoe --moe-cache-mib N --moe-cache-cpu-share F |
| External shards | Loads extra parts of a model from separate GGUF files, on any drive | --shard or --ple |
| Web UI | Chat groups, a markdown editor, settings in a sidebar footer, a phone layout | Open the server’s address in a browser |
The MoE cache is the one to try if a big model doesn’t fit on your GPU. --moe-cache-mib sets aside GPU memory for experts kept in system memory, and --moe-cache-cpu-share lets the processor work through some cache misses instead of copying them over. In the pull request that added the CPU share, its author’s test of Qwen3.8-Flash-Next ran at 24.2 tokens a second with a 6,464 MiB cache and 65.5 with a 19,392 MiB one, and it adds a script, scripts/moe-cache-tune.py, that tries settings on your machine.
What changed in llama-server
llama-server is the program in llama.cpp that serves a model over an OpenAI-style API, with a chat page in the browser. Version 1.0 closes several ways a web page or another computer could misuse it:
- Tools on the network need a key. Its built-in tools “can run shell commands,” as its code notes. If
--tools,--agent, MCP servers or the MCP proxy are on, the server listens on anything but a loopback address such as 127.0.0.1, and there’s no--api-key, it now refuses to start: “security: server tools, MCP servers or MCP proxy are enabled on a non-loopback address without an API key.” It suggests the fixes: “set –api-key, bind to 127.0.0.1, or pass –allow-unauthenticated-tools on a trusted network.” - Media URLs must be public. An image URL in a chat request can no longer reach your own computer, your local network, link-local cloud metadata addresses or the 100.64.0.0/10 range that Tailscale uses. The request fails with an error that reads “host is not a public address,” unless you start the server with
--media-allow-private-urls. The web UI’s MCP proxy reaches public addresses only, with, its pull request says, “no escape hatch.” - Origins are checked. Whenever
--cors-originsisn’t*, and--toolsand--agentset it tolocalhoston their own, browser requests from other origins get “Cross-origin request not allowed (see –cors-origins).”localhostnow means the server’s own page only, so a front end at, say,http://localhost:3000needs its own origin listed, port included. A DNS-rebinding check also refuses requests that reach a loopback address under any host name but a loopback one or an allowed origin’s. - Smaller fixes. Router mode’s child instances turn away browsers with “Browser requests must go through the router,”
HF_TOKENgoes only to Hugging Face’s download address, not to any URL a model comes from, and MCP tool calls stop when the client disconnects.
How serious were the holes? llama.cpp published no advisory, and its security policy doesn’t count the web UI, experimental features, the router or MCP as covered. The author of the tools change wrote: “None of those are actually vulnerability, but since people nitpick this too much, we now forced to hurt the UX a bit.” The media-URL change was different: reaching local addresses was by design, its pull request says, but “many deployments are now run directly on local computer, this started to become a real problem.”
One default hasn’t moved. Without tools, --cors-origins is still *, and with no API key the server only warns: “security: no API key is set and CORS allows all origins.”
What breaks
-dtand--defrag-tholdare gone. Remove them from commands, scripts and service files.- Tools on a network address without a key. Add
--api-key, or--allow-unauthenticated-toolsif you trust everyone on the network. - Images from your own network. Pictures served from a NAS or another machine on your network need
--media-allow-private-urls. - Front ends on other ports. List their origins in
--cors-origins, with the port. - Apps that link the library. Pull request 30565 removed
llama_set_warmup, “all 35 functions markedDEPRECATEDinllama.h” and thedefrag_tholdfield, and renamedllama_set_embd_state_srctollama_set_ctx_state_src.clip.his gone too, folded intomtmd. There is no migration guide.
How to update
| Where you got it | What you get on November 15 |
|---|---|
| Homebrew | brew upgrade llama.cpp installs 1.0.0, with ggml 1.0.0 |
| GitHub releases | The v1.0.0 tag has no binaries; download build b12359, the same code, or a newer nightly |
| winget | Nightly builds, such as b12395, which include everything in 1.0.0 |
| Docker | Pinned tags such as server-v1.0.0-b12359-b1b770d; the rolling server tag is newer |
| From source | git checkout v1.0.0, then build with -DLLAMA_BUILD_IS_DEV=OFF for a clean version string |
| Ollama | 0.40.16 bundles build b12177, from before 1.0, but already runs Ripple |
| LM Studio | No runtime built on 1.0 in its changelog yet |
Homebrew’s llama.cpp links to Homebrew’s own ggml package, also at 1.0.0. Ollama and LM Studio update their copies of llama.cpp on their own schedules.
Two models to check before you download
Inkling 2 Small isn’t in 1.0.0. Its support arrived in pull request 32217, merged late on November 14 and first in nightly build b12406, not the draft that Unsloth’s guide still tells you to build. Check which one a GGUF was made for before downloading a build of 171.7 GB or more. Our memo on Inkling 2 Small has the rest.
Kimi K3.1 Pro runs in 1.0.0, but not on a desk. The smallest file that stock llama.cpp loads, KarLexAI’s IQ1_S build, is 592.7 GB, and its card says “You need more memory than the file size, in GPU memory plus system RAM.”
Use it with a coding agent
llama-server listens on 127.0.0.1, port 9931, unless you tell it otherwise, so the new checks don’t get in the way of a coding agent on the same computer. Start a model with a short name, then point the agent at http://127.0.0.1:9931/v1:
llama-server -m model.gguf --jinja -c 32768 --alias local
In OpenCode, that’s a provider in opencode.json, as its provider docs describe, with local under models:
{
"$schema": "https://opencode.ai/config.json",
"provider": {
"llama.cpp": {
"npm": "@ai-sdk/openai-compatible",
"name": "llama-server (local)",
"options": { "baseURL": "http://127.0.0.1:9931/v1" },
"models": { "local": { "name": "My local model" } }
}
}
}
Enjoy runs OpenCode beside Claude Code, Codex and the other labs’ agents, and lists the models OpenCode has set up, local ones included. When another agent’s plan runs low, you can switch the conversation to your own model and keep going, and nothing leaves your computer. OpenCode 2 finds models from Ollama, LM Studio or vLLM on its own; llama-server needs the entry above.
What ggml-org hasn’t said
- How often releases will come. It hasn’t given a schedule for 1.0.x or 1.1, and nightly builds continue.
- When winget and the llama.app installer switch. Its versioning page says both publish “the nightly builds instead of the official releases,” and leaves open whether to “switch to official releases?”
- A list of breaking changes. The release notes don’t have a section for them; a contributor asked for one on November 15.
9.Questions.
Is llama.cpp 1.0 out?
- Yes. ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on November 13, 2026. It’s the same code as nightly build b12359, which carries the downloadable binaries, and Homebrew’s llama.cpp formula is already at 1.0.0.
Should I update llama.cpp?
- If you run llama-server where other computers or web pages can reach it, yes. Version 1.0 refuses tools on a network address without an API key, stops media URLs from reaching private addresses, and checks browser origins more strictly. If you only use it on 127.0.0.1, update for the new models and speedups when it suits you.
Why won’t llama-server start with “server tools, MCP servers or MCP proxy are enabled on a non-loopback address without an API key”?
- Because version 1.0 refuses to expose its tools, which can run shell commands, to the network without a key. Set --api-key, bind to 127.0.0.1 with --host, or on a network you trust pass --allow-unauthenticated-tools.
Will llama.cpp 1.0 break my commands or scripts?
- Mostly not. The -dt and --defrag-thold option is gone, so take it out. Image URLs that point to your own computer or local network now fail with “host is not a public address” unless you add --media-allow-private-urls. Apps that link the library lose 35 deprecated functions.
Does llama.cpp 1.0 run Ripple-65B, Kimi K3.1 Pro and Inkling 2?
- Ripple-65B and Kimi K3.1 Pro, yes, though the smallest Kimi K3.1 Pro file it loads is 592.7 GB. Inkling 2 support landed a day later, in nightly build b12406 on November 14, so 1.0.0 doesn’t have it.
Do Ollama and LM Studio use llama.cpp 1.0?
- Not yet. Ollama 0.40.16, from November 9, bundles llama.cpp build b12177 and already runs Ripple. LM Studio’s changelog shows no runtime built on 1.0 by November 15. Both ship their own copies of llama.cpp on their own schedule.