llama.cpp reaches v1.0.0 with API cleanup and new model support
The llama.cpp project has released version 1.0.0, removing deprecated APIs, adding support for Kimi-K3.1 and Ripple models, and hardening server security.
The maintainers of llama.cpp have announced the release of version 1.0.0, marking a significant milestone for the popular local large language model inference engine. Published on November 13, 2026, this major update introduces substantial breaking changes to the public application programming interface while expanding support for new model architectures like Kimi-K3.1-Pro and Ripple. The release also focuses heavily on security hardening for the built-in server and performance optimizations within the underlying ggml library.
What happened
This release represents a maturation point for the project, characterized by the removal of technical debt and the stabilization of core interfaces. The developers removed thirty-five deprecated functions from the public header file llama.h, including llama_set_warmup and various threshold parameters. This cleanup aims to simplify the codebase for developers integrating the library into their own applications. Alongside these removals, the NextN speculative decoding API was promoted from the experimental llama-ext.h header to the main llama.h interface, signaling its readiness for production use.
Significant attention was paid to multi-modality and new model architectures. The update adds native support for three new text-based architectures: Kimi-K3.1-Pro, which replaces full-attention layers with DSA; Ripple-65B-Preview, a dense 65-billion parameter model; and xing4_0, which utilizes a mix of hybrid attention and mixture-of-experts techniques. Additionally, the system now supports drama-3 text-to-speech capabilities and vision processing for MageVL-Cowork and Kimi-K3.1 models. These additions broaden the range of tasks that can be performed locally without relying on external cloud services.
Security and server stability were also major focal points. The llama-server component received patches to address server-side request forgery, cross-origin resource sharing issues, and DNS rebinding vulnerabilities. The team introduced opt-in flags for allowing unauthenticated tools and private media URLs, giving administrators finer control over access. Furthermore, the release includes a redesigned web user interface with features such as chat groups, a markdown editor, and improved mobile responsiveness, making the default front-end more usable for daily interactions.
How it works
Under the hood, the release updates the foundational ggml library to version 1.0.0, which prunes legacy operations and introduces configurable FP8 support. FP8, or 8-bit floating point, allows for more efficient memory usage during inference, particularly for key-value caches. The system now supports OCP FP8 KV caches with optional static per-tensor scales, enabling faster processing of long contexts while reducing memory bandwidth requirements. This is complemented by a host-memory mixture-of-experts cache that allows users to configure how much of the expert data resides in CPU memory, reducing wait times when cached experts are accessed.
The multi-modal infrastructure was consolidated into a unified mtmd module, removing the older clip.h dependency. This change allows different subcontexts, such as vision and audio processors, to share compute buffers, saving significant memory resources. For audio generation, the system now computes log-mel spectrograms directly within the computation graph rather than spawning separate threads for each request. This architectural shift reduces overhead and improves latency for real-time audio tasks. The server also gained an OpenAI-compatible endpoint for audio speech generation, standardizing how applications interact with local text-to-speech models.
Key details
- The public API in
llama.hremoved 35 deprecated functions and promoted the NextN speculative decoding API. - New model support includes Kimi-K3.1-Pro, Ripple-65B-Preview, xing4_0, and drama-3 text-to-speech.
- Server security was hardened against SSRF, CORS abuse, and DNS rebinding attacks with new opt-in flags.
- FP8 key-value caches are now supported, offering memory efficiency for long-context inference.
- The web UI was redesigned with chat groups, a live-preview markdown editor, and better mobile layout.
- The underlying ggml library reached v1.0.0, adding FP8 support and optimizing backend performance for CUDA, Metal, and Vulkan.
Why it matters
For software engineers building products on top of local large language models, this release signals a shift toward greater stability and predictability. The removal of deprecated APIs means that integrations built against older versions will require updates, but the resulting codebase will be cleaner and easier to maintain. The promotion of the NextN API suggests that speculative decoding is now considered a standard feature, allowing developers to implement faster inference strategies with confidence in the interface's longevity.
The security improvements in llama-server are critical for teams deploying these models in networked environments. Previous vulnerabilities could have allowed malicious actors to exploit the server for unauthorized access or data exfiltration. By addressing these issues and providing granular controls over tool authentication and media access, the release makes it safer to expose local models to internal networks or trusted users. This is particularly important for enterprises looking to adopt local AI without compromising their security posture.
Performance optimizations such as FP8 caches and shared compute buffers for multi-modal tasks enable more complex workflows on consumer hardware. Developers can now run vision and audio models alongside text generation with lower memory overhead, making it feasible to build richer multi-modal applications on devices with limited resources. This democratizes access to advanced AI capabilities, allowing smaller teams to prototype and deploy sophisticated systems without expensive cloud infrastructure.
What you can do
- Review your integration code to replace calls to the 35 removed deprecated functions in
llama.h. - Update your deployment scripts to include the new security flags for
llama-serverif exposing it to a network. - Test the new FP8 KV cache settings to evaluate memory savings and performance gains for your specific models.
- Explore the new
/v1/audio/speechendpoint if your application requires local text-to-speech capabilities. - Upgrade the ggml backend dependencies to v1.0.0 to benefit from the latest performance optimizations across GPU platforms.
- Experiment with the new web UI features like chat groups to improve user experience in internal tools.



