llama.cpp v1.0 Is Here: What Changed and How to Upgrade Your Self-Hosted Stack
After years of being the de-facto standard for running LLMs on anything with a CPU and some patience, llama.cpp hit v1.0.0 on November 13, 2026. It's the first stable major release — 135,000+ GitHub stars, the engine underneath Ollama, LM Studio, and half the "local AI" tutorials on the internet — and it ships alongside ggml v1.0.0 (the tensor library it's built on, released the same day).
The headline: this is a breaking release. The llama.h public API got pruned hard, deprecated functions are gone, and some of the new features change the way you configure your server. If you're running llama.cpp in a Proxmox LXC, a Docker container, or bare metal, you should know what changed before you pull the update — because a lot of the "how to run a local LLM" guides written in the last two years will now be subtly wrong.
Here's the honest, self-hoster-focused tour: what's new, what breaks, and the exact upgrade path.
Why v1.0 Matters (and Why It's Breaking)
llama.cpp has been on a 0.x version forever, which quietly meant "the API can change under you." v1.0.0 is the project saying the surface is now stable enough to commit to — but getting there required removing the deprecated public API and promoting the things that had been living in experimental headers. That's the "good pain": one-time breakage now, in exchange for a stable target afterward.
The practical consequences:
- 35 deprecated functions are gone from
llama.h, plusllama_set_warmupand thedefrag_thold/--defrag-tholdoption. - The NextN / speculative-decoding API was promoted from
llama-ext.hinto the main header — andllama_set_embd_state_srcwas renamed tollama_set_ctx_state_src. - ggml's RPC protocol bumped to v9, which rejects incompatible peers. If you run distributed inference across multiple nodes, they all need to be on the same version or they'll refuse to talk.
If you build llama.cpp from source and only use the official llama-server binary, you're mostly insulated from the C API churn — but if you compile against llama.h (or use a binding that does), expect a recompile and possibly a small port.
What's Actually New (the Self-Hoster's Shortlist)
The release notes are long; here are the changes that genuinely move the needle for a homelab:
1. FP8 KV caches (-ctk/-ctv f8_e4m3)
KV caches are the hidden memory hog in long-context inference. v1.0 adds FP8 KV cache support with optional static per-tensor scales. In plain terms: you can hold roughly twice the context in the same VRAM (or RAM) at a small quality cost. If you've been running out of memory on long documents or big agent contexts, this is the single biggest win in the release.
# halve KV-cache memory with FP8 cache types
llama-server -m model.gguf -ctk f8_e4m3 -ctv f8_e4m3
2. MoE expert cache with a configurable CPU share
Mixture-of-Experts models (the big frontier MoEs — think Mixtral-style and larger) only activate a fraction of their parameters per token, but naively they still load all experts. v1.0 adds a host-memory expert cache where cache misses can be computed on the CPU and copied back asynchronously — so you can run an MoE that's larger than your GPU VRAM and let the CPU soak up the expert lookups it can handle. Two new knobs in llama_context_params:
moe_cache_size— how many experts to keep residentmoe_cache_cpu_share— what fraction of misses the CPU computes
This is the "run a 100B+ MoE on a single consumer GPU" feature, and it's the closest thing to free VRAM this release.
3. External PLE / engram shards (--shard / --ple)
You can now load separate GGUF shards for "PLE"/engram components via --shard and --ple, instead of one monolithic file. For very large models that don't fit in memory all at once, this means finer control over what stays resident and what streams from disk.
4. New model architectures
- Kimi-K3.1-Pro — K3 with the full-attention MLA layers replaced by DSA.
- Ripple-65B-Preview — a new dense 65B architecture (no new ops required).
- xing4_0 — mHC + MLA + MoE + MTP.
- drama-3 TTS (fishaudio) — text-to-speech in llama.cpp.
- MageVL-Cowork and Kimi-K3.1 vision — new multimodal/vision support.
The TTS one is quietly notable for self-hosters: it means llama-server can now also be your speech synthesis endpoint, not just your LLM.
5. A hardened llama-server
This is the part I care most about for a homelab that's exposed to more than localhost:
- SSRF protection, CORS handling, and DNS rebinding defenses — three of the classic ways a local inference server gets abused or exfiltrated.
- MCP tool cancellation — so a runaway tool call doesn't hang the server.
- A new OpenAI-compatible
/v1/audio/speechendpoint (and atemperatureoption for/tts). - A redesigned web UI with chat groups and a markdown editor.
If you've been nervous about putting llama.cpp's server on your LAN (or behind a reverse proxy), the SSRF/CORS/DNS-rebinding hardening is the reason to upgrade even if you don't care about FP8.
ggml v1.0.0: the Engine Underneath
Released the same day, ggml v1.0.0 is mostly invisible to end users but matters for anyone doing custom work. The short version:
- Pruned the deprecated API and removed legacy
DIAG_MASKops — another one-time breakage for the "commit to stability" goal. - Configurable YaRN truncation (rotation correction-bound rounding) via new
ggml_rope_set_truncate/ggml_rope_get_truncate. - FP8 support added.
- Broad backend performance work: few-row quantized matmuls and MoE fusion on CUDA/Metal, a Vulkan prefill FlashAttention path, a device-side MoE scheduler on SYCL, decomposed Adreno FlashAttention in OpenCL, new WebGPU kernels, expanded Hexagon op coverage, and scheduler improvements that stop uploading unused inputs.
The one user-facing gotcha: the RPC protocol bump to v9 (faster loading, graceful failure on unreachable endpoints). Distributed setups across multiple machines must match versions.
How to Upgrade (Step by Step)
Option A — Build from source (recommended for Proxmox / bare metal)
cd ~/llama.cpp
git fetch && git checkout v1.0.0
# CPU-only build
cmake -B build && cmake --build build --config Release -j$(nproc)
# CUDA build (NVIDIA GPU)
cmake -B build -DGGML_CUDA=ON && cmake --build build --config Release -j$(nproc)
# Metal build (Apple Silicon Mac)
cmake -B build -DGGML_METAL=ON && cmake --build build --config Release -j$(nproc)
Then verify:
./build/bin/llama-server --version
If you use a language binding (Python llama-cpp-python, Node bindings, etc.), rebuild it against the new headers — the removed functions will cause compile errors that tell you exactly what to update.
Option B — Docker
The official image tags move with releases. Pin to the v1.0.0 tag rather than latest so the breaking change is a deliberate act, not a surprise:
docker pull ghcr.io/ggml-org/llama.cpp:v1.0.0
# or, for CUDA:
docker pull ghcr.io/ggml-org/llama.cpp:v1.0.0-cuda
Option C — Ollama / LM Studio
If you use Ollama or LM Studio, you don't need to do anything right now — they bundle and pin llama.cpp internally and will roll v1.0 support into their own releases. Your models keep working. The FP8 KV cache and MoE expert-cache benefits will arrive when they upgrade their runtime, not when you do.
What Breaks (a Concrete Checklist)
Going into v1.0.0, these are the failure modes to watch for:
- C API consumers won't compile. If you build anything against
llama.h, the 35 removed functions + thellama_set_embd_state_src→llama_set_ctx_state_srcrename will surface as errors. It's a rename-and-recompile, not a redesign. --defrag-tholdis gone. Any scripts or systemd units passing it will fail. Remove the flag.llama_set_warmupis removed. If your warmup logic called it, drop the call.- Distributed RPC peers must match. If you shard across machines with the RPC backend, upgrade all of them together or they'll refuse to connect (protocol v9).
- The web UI changed. Chat groups + markdown editor are new; if you bookmarked or scripted against the old UI's structure, re-check it.
None of this is catastrophic — it's exactly what a "1.0" is supposed to do: break old cruft once, then hold still. But if you have a systemd unit or cron job that boots llama-server with old flags, check it before you reboot the box.
Should You Upgrade Now?
- If you self-host llama.cpp directly — yes, on your schedule. The server hardening (SSRF/CORS/DNS rebinding) alone is worth it if your server is reachable beyond localhost, and FP8 KV caches are a genuine memory win for long-context work.
- If you're on Ollama/LM Studio — wait for their release; you get the benefits when they ship.
- If you run MoE models larger than your GPU — this is the release you've been waiting for. The MoE expert cache with CPU share is the headline feature.
- If you build against the C API — budget an hour for the recompile/rename, but do it soon; you don't want to fall multiple breaking versions behind.
The quiet significance here isn't any single feature — it's that llama.cpp committed to stability. A 1.0 means the tooling built on top of it (bindings, runtimes, orchestration layers) now has a fixed target. For the self-hosted AI ecosystem, that's a big deal: fewer "it worked yesterday and broke today" moments, more durable tutorials, and a saner upgrade story going forward.
Sources: llama.cpp v1.0.0 release notes, ggml v1.0.0 release notes, and the Hugging Face llama-cpp-v1 blog post. Star count and release dates checked live at the time of writing.