I moved a CPU box from the previous stable tag to llama.cpp 1.0 the day it was tagged. What broke was a removed flag, a new libllama soname, 36 C API symbols gone, and a llama-server that now refuses setups the old build only warned about. Each claim below sits next to the command that proved it and the PR number from the release notes, so you can rerun any of it before your own upgrade.
I ran all of it in November 2026 on an Ubuntu 26.04.1 LTS server VM called box1 (4 vCPUs on AMD EPYC, 16 GB RAM, no GPU) with the prebuilt v1.0.0 CPU archive next to the v0.9.0 build it replaces, and every prompt below comes from that box.
Finding the build behind the llama.cpp 1.0 tag
The v1.0.0 release on GitHub carries no binaries, only nightly-tag.txt, which names the nightly build the tag points at, so read that file first from a working directory such as ~/llama:
TAG=$(curl -fsSL https://github.com/ggml-org/llama.cpp/releases/download/v1.0.0/nightly-tag.txt) # pin: this post tests the v1.0.0 tag #https://github.com/ggml-org/llama.cpp/releases/tag/v1.0.0
echo $TAG
On release day it pointed at build b12359:
b12359
Download, checksum (compare with the digest GitHub shows on the b12359 release page) and unpack into its own directory:
curl -fsSLO https://github.com/ggml-org/llama.cpp/releases/download/$TAG/llama-$TAG-bin-ubuntu-x64.tar.gz
sha256sum llama-$TAG-bin-ubuntu-x64.tar.gz
tar -xzf llama-$TAG-bin-ubuntu-x64.tar.gz && cd llama-$TAG
./llama-server --version
On a minimal Ubuntu image, --version fails first with a missing libgomp.so.1. Both builds need it, so run sudo apt-get install -y libgomp1 (plus jq for the JSON filters later), and the binary then reports:
version: 1.0.0-dev (build 12359, commit b1b770d90)
built with GNU 11.4.0 for Linux x86_64
The -dev suffix is expected, because the nightly pipeline built it. To prove which release you run, match the build number and short commit, not the version string. The whole sequence on box1, failure included:

For comparisons I resolved v0.9.0 the same way (b12110) into ~/llama/b12110. Keep each build in its own directory, because every archive ships dozens of shared libraries. The test model was Gemma 3 1B, which loads in a second on CPU:
curl -fsSL --create-dirs -o ~/models/gemma-3-1b-it-Q4_K_M.gguf https://huggingface.co/ggml-org/gemma-3-1b-it-GGUF/resolve/main/gemma-3-1b-it-Q4_K_M.gguf
For a first install from scratch, see our guide to running a local LLM with llama.cpp. This post only covers the move to 1.0.
llama.cpp 1.0 breaking changes: the removed –defrag-thold flag
PR #30565 deleted -dt/--defrag-thold and the defrag_thold context field. This is the removal most likely to break a systemd unit or Docker command:
./llama-server -m ~/models/gemma-3-1b-it-Q4_K_M.gguf --defrag-thold 0.1; echo exit=$?
The 1.0 build exits before loading the model:
error: invalid argument: --defrag-thold
exit=1
The b12110 build only logged W DEPRECATED: --defrag-thold is deprecated and no longer necessary to specify and kept going. That warning is easy to miss in a journal, so a unit that ran fine for months will crash loop on its first restart after the upgrade. Oddly, LLAMA_ARG_DEFRAG_THOLD=0.1 is silently ignored and the server starts normally, which means a compose file that sets it through the environment survives while one that passes the flag dies.

Delete the flag, because there’s nothing to replace it with. The old warning already said it was no longer necessary, so removing it changes nothing about how the cache behaves. Grep units, compose files and wrapper scripts for -dt and defrag before you swap binaries, since the failure only shows up at the next restart.
Testing the FP8 KV cache on a CPU
PR #28392 adds the OCP FP8 KV cache, -ctk f8_e4m3 -ctv f8_e4m3. I started a 32K context with each cache type and read the sizes from the verbose log:
for t in f16 q8_0 f8_e4m3; do echo "== $t"; timeout 15 ./llama-server -m ~/models/gemma-3-1b-it-Q4_K_M.gguf -c 32768 -ctk $t -ctv $t -fa on -lv 4 2>&1 | grep "llama_kv_cache: size"; done
Gemma 3 prints a full-context and a sliding-window line per type, so the totals add both:
| Cache type | Full-context part | Sliding-window part | Total at 32K |
|---|---|---|---|
| f16 (default) | 128.00 MiB | 55.00 MiB | 183 MiB |
| q8_0 | 68.00 MiB | 29.22 MiB | 97 MiB |
| f8_e4m3 | 64.00 MiB | 27.50 MiB | 91.5 MiB |
FP8 halves f16 exactly and lands about 6% under q8_0. On a 70B model at long context, that saving can decide which GPU you need for local LLMs. The old build rejects the type outright:
error while handling argument "-ctk": Unsupported cache type: f8_e4m3
Both runs together, the three sizes from 1.0 and the refusal from b12110:

Then speed, with llama-bench over all nine K and V combinations (over ten minutes on this box):
./llama-bench -m ~/models/gemma-3-1b-it-Q4_K_M.gguf -fa 1 -ctk f16,q8_0,f8_e4m3 -ctv f16,q8_0,f8_e4m3 -p 512 -n 128 -r 3
Trimmed to the rows that matter, in tokens per second:
| K cache | V cache | pp512 t/s | tg128 t/s |
|---|---|---|---|
| f16 | f16 | 73.44 | 22.06 |
| q8_0 | q8_0 | 70.45 | 23.69 |
| f8_e4m3 | f16 | 64.39 | 22.53 |
| f8_e4m3 | q8_0 | 61.94 | 22.24 |
| f16 | f8_e4m3 | 22.18 | 19.84 |
| f8_e4m3 | f8_e4m3 | 21.43 | 19.55 |
Read the table this way:
- FP8 V cuts prompt processing 2.9x to 3.3x
- FP8 K costs only about 12%
- Generation drops 10 to 19% once V is FP8
q8_0/q8_0is about 4% behind f16 on prompts

On a CPU box I’d use q8_0 for both K and V, because it’s under 6 MiB bigger than FP8 here and doesn’t pay the 3x prompt penalty. FP8 targets GPUs with native FP8 math (CUDA support is #28413), and I measured only one EPYC VM with a 1B model and no GPU, so treat these numbers as a CPU warning, not a verdict on FP8. Our Ollama vs vLLM vs llama.cpp benchmark shows the engine on real GPU hardware.
New llama-server flags in llama.cpp 1.0
Diffing llama-server --help between b12110 and b12359 shows six new flags and one removal. Two of the new ones are opt-outs from security hardening, which tells you what now gets refused by default:
| Flag | What it changes | PR |
|---|---|---|
--allow-unauthenticated-tools | Lets --tools, MCP servers or the MCP proxy run on a non-loopback address without --api-key. Without it, the server refuses to start. | #32113, #32115 |
--media-allow-private-urls | Lets image_url and other media URLs point at private or local addresses. The default now allows public addresses only. | #32113, #32115 |
--moe-cache-cpu-share F | Share of MoE cache misses the CPU computes from host experts instead of uploading them to the GPU (0 to 1, default 0). | #30966, #31233, #31888 |
--shard / --ple | Loads extra GGUF shards (PLE and engram weights) from any path, even another drive. | #31185 |
--tts-lang, --tts-voice | Language and voice for audio generation, alongside the new OpenAI-style /v1/audio/speech endpoint. | Flags seen in --help, no PR in the release body; endpoint #31763 |
-dt, --defrag-thold | Removed. Passing it is now a fatal argument error. | #30565 |
Changes with no new flag, which you’ll only notice when a client gets refused:
--cors-originsfixed, includinglocalhostmeaning same-origin only (#32098)- DNS rebinding requests blocked (#32098)
- Web UI CORS proxy limited to public addresses, no opt-out (#32115)
- Browsers kept away from router child instances (#32120)
- MCP tool calls cancelled on client disconnect (#32025)
I didn’t use a cross-origin front end or the CORS proxy on box1, so these went untested. If you point the web UI’s MCP proxy at a server on localhost or your LAN, expect it to stop working, because #32115 gives the proxy no opt-out. The llama.cpp cheat sheet lists the matching LLAMA_ARG_* environment variables for most server flags.
Other server changes, not tested here, as the release body words them:
- Reference-counted context checkpoints instead of deep copies (#30137)
- Parent slots visited before child slots for shared-prefix requests (#30779)
- Realtime API sends error events instead of crashing, rejects non-object session updates, fixes the output lang (#31632, #31714, #31982)
- Top-level
response_format.schemaform accepted, Responsestext.formatJSON schemas converted (#28697, #26145)
None of these change a command line, so they need no action during the upgrade.
The tools refusal on 0.0.0.0
This one will break the most home labs: bind to 0.0.0.0 for the web UI, then switch on tools. The old build warned; 1.0 exits:
./llama-server -m ~/models/gemma-3-1b-it-Q4_K_M.gguf --host 0.0.0.0 --tools all
The error names all three ways out:
E srv server_setup: security: server tools, MCP servers or MCP proxy are enabled on a non-loopback address without an API key
E srv server_setup: set --api-key, bind to 127.0.0.1, or pass --allow-unauthenticated-tools on a trusted network
What each one did on box1:
--api-key: starts, listening on0.0.0.0on the configured port- Loopback with no key: starts
--allow-unauthenticated-tools: starts, but the tool list includesexec_shell_commandandwrite_file, so keep it to a network you fully own

An API key is the right fix for almost every LAN setup, because it keeps the web UI reachable from other machines and puts the shell tools behind a secret. Loopback suits a single-user laptop, but then nothing else on the network can reach the server. So who should pass the opt-out flag? Only a lab on an isolated network where every host is yours.
Private image URLs are blocked now
The SSRF fix in #32115 also stops local vision pipelines. I served a solid red red.png on loopback and started SmolVLM:
python3 -m http.server 8000 --bind 127.0.0.1 &
./llama-server -hf ggml-org/SmolVLM-256M-Instruct-GGUF --port 8081 &
Then sent an OpenAI-style image message once the server was listening:
REQ='{"messages":[{"role":"user","content":[{"type":"text","text":"What colour is this image? One word."},{"type":"image_url","image_url":{"url":"http://127.0.0.1:8000/red.png"}}]}],"max_tokens":10}'
curl -s http://127.0.0.1:8081/v1/chat/completions -H 'Content-Type: application/json' -d "$REQ" | jq -c '.error // .choices[0].message.content'
Without the new flag the server returns HTTP 400 instead of fetching:
{"code":400,"message":"host is not a public address: 127.0.0.1","type":"invalid_request_error"}
Restarted with --media-allow-private-urls, the same request returned " The image is a red gradient background. The gradient":

If images live on an internal MinIO, a NAS or another container, your client now gets that 400 instead of an answer. The flag fixes it, but it reopens the SSRF hole for every caller of that server, so set it only where every client is trusted.
Testing whether HF_TOKEN stays on huggingface.co
PR #31730 stops HF_TOKEN riding along on -mu/--model-url downloads from other servers. The token is now kept only when the URL’s scheme, host and port match the Hugging Face endpoint and it starts with hf_. A custom MODEL_ENDPOINT is still trusted. My test went like this.
- Tiny Python server on 127.0.0.1:8000 logging the
Authorizationheader - Both builds run with
HF_TOKEN=hf_fakefakefakefakeand-mu http://127.0.0.1:8000/models/tiny.gguf
Every HEAD and GET from both builds logged Authorization: <none>, so the old leak didn’t reproduce on plain HTTP loopback and I’m taking the PR’s description on trust. In practice: if a private mirror relied on receiving HF_TOKEN, it stops on 1.0 unless that mirror is your MODEL_ENDPOINT.
Retesting –sleep-idle-seconds
--sleep-idle-seconds isn’t new, since b12110 has it. What 1.0 changes is keeping the server awake while a request still uses its state (#32133), which fixes a read of freed memory when a request arrived as it went to sleep. I ran it with an FP8 cache and a 15 second timeout:
./llama-server -m ~/models/gemma-3-1b-it-Q4_K_M.gguf -c 8192 -ctk f8_e4m3 -ctv f8_e4m3 -fa on --sleep-idle-seconds 15 --port 8080 > ~/chat.log 2>&1 &
After one chat request and a wait past the timeout, curl -s 127.0.0.1:8080/props reported "is_sleeping": true and "build_info": "b12359-b1b770d90", and the next chat request woke it with HTTP 200 in 1.308952s:

Chat responses also carry "system_fingerprint": "b12359-b1b770d90", a handy way to confirm the build from a client. Wake time grows with model size, since it includes the reload.
Upgrading to llama.cpp 1.0: libllama C API removals
This matters if you use bindings (llama-cpp-python, Go or Rust wrappers) or link libllama yourself. From ~/llama:
for d in b12110/llama-b12110 llama-b12359; do readelf -d $d/libllama.so | grep SONAME; nm -D --defined-only $d/libllama.so | grep -c " T llama_"; done
The soname moved to .so.1, so old binaries won’t load it, and exported llama_* functions fell from 255 to 237:
0x000000000000000e (SONAME) Library soname: [libllama.so.0]
255
0x000000000000000e (SONAME) Library soname: [libllama.so.1]
237
A comm of the sorted lists shows 36 removals, matching the release notes: 35 deprecated functions plus llama_set_warmup. The net drop is smaller because the NextN speculative API moved into llama.h and llama_set_embd_state_src became llama_set_ctx_state_src:

The swaps, each checked against the new exports:
llama_load_model_from_file,llama_free_model: usellama_model_load_from_file,llama_model_freellama_new_context_with_model: usellama_init_from_modelllama_token_*andllama_n_vocab: use the matchingllama_vocab_*call (llama_vocab_bos,llama_vocab_eos,llama_vocab_n_tokens) onllama_model_get_vocab()llama_token_cls,llama_vocab_cls: usellama_vocab_bosllama_add_bos_token,llama_add_eos_token: usellama_vocab_get_add_bos,llama_vocab_get_add_eosllama_sampler_init_grammar_lazy: usellama_sampler_init_grammar_lazy_patternsllama_n_ctx_train,llama_n_embd,llama_n_layer,llama_n_head: use thellama_model_n_*versionsllama_*_state_data,llama_*_session_file,llama_get_state_size: usellama_state_*llama_set_warmup: no replacement, run your own warmup pass
An old binding fails with an undefined symbol at build or load time, not a silent behaviour change, so you’ll find out on the first start rather than in wrong output. Most of these names have had replacements for many releases, which means a binding that’s still maintained should already be clean. The clip.h header is gone too (#31865), which only matters if you built against the multimodal internals instead of mtmd.
Stable tags or nightlies: the 1.0 upgrade checklist
The ggml-org releases and versioning discussion sets two tracks for llama.cpp. Official releases use semantic vX.Y.Z tags, nightly builds are tagged b[NUM] by commit number, and a release tag is only created when llama.cpp’s internal ggml copy is identical to the released ggml. That last rule is what makes a stable tag safe to build against a system-wide ggml. On a box that runs llama.cpp as a service:
- Pin a stable tag, resolve it through
nightly-tag.txt, and record both the tag and the build number - Unpack into a new directory, never over the old one
- Install
libgomp1on minimal Ubuntu images - Remove
-dt/--defrag-tholdfrom units and scripts - Add
--api-keyif tools or MCP run on a non-loopback address - Check vision clients that pass private image URLs
- Upgrade bindings to a release built against
libllama.so.1 - Confirm the build from a client through
system_fingerprint
If llama.cpp sits behind a service other people depend on, pin vX.Y.Z and move when a new tag lands. If you chase new model architectures the week they ship, stay on bNNNN, but never treat two nightlies as interchangeable, because the C API can change between any two of them. Whichever track you follow, test your bindings against each new build before it reaches the service.