llama.cpp Hits Version 1.0.0: Faster on NVIDIA, Intel, and ARM, SSRF Holes Closed
The llama.cpp engine, which many people use to run neural networks locally, has reached version 1.0.0 (the previous version was 0.9.0, build b12359). The release includes speedups for NVIDIA, Intel graphics, and ARM processors, and it also closes SSRF holes that could be used to make the server send requests to other addresses, such as the internal network. Old examples and APIs have been removed. If llama-server is exposed to a network, update it. If you build from source, check for deprecated calls first.
AI-processed from TG: @cgevent (Метаверсище и ИИще); edited by Hamidun News
If you have run a language model on your laptop, a mini server, or a phone, there is a good chance llama.cpp or a tool built on it was running under the hood. The project and its underlying ggml library are now at version 1.0.0 (build b12359), up from 0.9.0. The version number itself speeds nothing up, but a large set of changes shipped with it. Obsolete code was removed, optimizations were added for various hardware, and SSRF vulnerabilities were closed. SSRF is an attack in which a server is tricked into sending requests to addresses chosen by an attacker, such as a company's internal network.
What changed in versioning and code
llama.cpp (release #32190) and ggml both shipped as v1.0.0 at the same time. The authors used the occasion to clear legacy code from the llama and ggml codebases: they removed outdated examples and part of the old APIs.
For those who simply download prebuilt binaries, this is almost invisible. But if you build the project from source, use ggml as a library, or rely on examples that sat in the repository for years, read the list of changes and test your builds before updating. ggml-org has a separate discussion of its release policy. What exactly it promises about compatibility is not visible from the source text, so look to the primary source for guarantees.
Speedups for different hardware
The bulk of the release consists of low-level optimizations:
- CUDA (NVIDIA). Few-row MMQ kernels were added for small batches of 2 to 24. This is the case where a server handles only a few requests at once rather than hundreds. For models in the NVFP4 format, prefill (the initial processing of the prompt) was optimized: the w_s scales are now fused into the final step of the MMQ kernel.
- Vulkan. A FlashAttention kernel for the prefill stage was added to the backend for Intel graphics. Vulkan is often chosen where CUDA is unavailable, for example on the integrated graphics of laptops.
- ARM and RISC-V. 4x4 simd_gemm tiles for ARM NEON now handle matrix multiplication on smartphone and compact board processors. The GATED_DELTA_NET operation is now routed to SpacemiT RVV (RISC-V) vector kernels.
- Qualcomm Hexagon. Q5_0 and Q5_1 quantization was added, along with a Windows ARM64 host build.
The source gives no speed-gain figures, so promising a specific percentage improvement is not possible. The effect depends on the model, the quantization, and the hardware. It is more sensible to run your typical request on the old and new builds and compare time to first token and overall throughput.
Server: MCP and security
For anyone using llama-server as the basis for agents, the most practical change is this: if the client drops the connection or the timeout expires, the server now cancels MCP tool calls. This matters for tools with side effects, such as writing a file or querying an external service. Unfinished work should not keep running when no one is waiting for the response.
SSRF vulnerabilities in the CORS proxy are closed. Imagine that someone could ask the proxy on your server to open an address of an internal admin service that is unreachable from outside. The server would go there itself and return the response. After the fix, this trick should no longer work. In addition, access to tools from non-local addresses has been tightened. If you exposed llama-server to a shared network or to the internet, that is a serious reason to update. Even after updating, however, do not leave it open without authentication and a reverse proxy.
Separately, the conversion of the Qwen3.5 model and of the multimodal projectors for Qwen-VL, HunyuanVL, and MiniCPM-V was fixed. If you had problems converting these models to GGUF, it is worth trying again.
What to do in practice
- Read the release notes on GitHub and check whether you use any removed examples or APIs.
- If you run llama-server with MCP tools or with access from beyond localhost, update first because of the security fixes.
- Compare speed on your own hardware before and after: the source contains no benchmarks, and the gain may vary widely.
Release v1.0.0 is more a sign of the project's maturity than a revolution. One of the core tools of local AI gets a clean start and patches where they were needed.
- Nov 21, 2026A 16 GB Mac mini instead of a paid API: Qwen analyzes food photos in 20–40 seconds
- Nov 21, 2026Cloudflare Open-Sources Kitesurf, a Browser for AI Agents That Runs Without Chromium
- Nov 21, 2026Datasette Agent Runs on Local Qwen 27B Without the Cloud via LM Studio and Ollama
- Nov 21, 2026An Engineer Ran a 125-Billion-Parameter Model on a Home PC at 22–25 Tokens per Second
Key Questions Answered
What changed in versioning and code?
llama.cpp (release #32190) and ggml both shipped as v1.0.0 at the same time. The authors used the occasion to clear legacy code from the llama and ggml codebases: they removed outdated examples and part of the old APIs. For those who simply download prebuilt binaries, this is almost invisible. But if you build the project from source, use ggml as a library, or rely on examples that sat in the repository for years, read the list of changes and test your builds before updating.
What are the key details on Speedups for different hardware?
The bulk of the release consists of low-level optimizations: CUDA (NVIDIA). Few-row MMQ kernels were added for small batches of 2 to 24. This is the case where a server handles only a few requests at once rather than hundreds. For models in the NVFP4 format, prefill (the initial processing of the prompt) was optimized: the w_s scales are now fused into the final step of the MMQ kernel. Vulkan. A FlashAttention kernel for the prefill stage was added to the backend for Intel graphics.
What are the key details on Server: MCP and security?
For anyone using llama-server as the basis for agents, the most practical change is this: if the client drops the connection or the timeout expires, the server now cancels MCP tool calls. This matters for tools with side effects, such as writing a file or querying an external service. Unfinished work should not keep running when no one is waiting for the response. SSRF vulnerabilities in the CORS proxy are closed.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.