Models & Labs
llama.cpp adds Inkling 2 text and vision support
llama.cpp Releaseshigh confidence
Why it matters
- →Enables local inference of Meta's new Inkling 2 MoE architecture.
- →Adds multimodal vision capabilities via mtmd for image understanding.
- →Supports complex quantization schemes like NVFP4 and MXFP8 for efficient deployment.
llama.cpp has released version b12406, adding support for Meta's Inkling 2 model architecture. The update includes text inference capabilities and multimodal vision support through the mtmd module, enabling image understanding alongside text generation. Technical implementation details include handling of NVFP4 and MXFP8 quantization schemes for sparse experts, with tests validating bit-for-bit accuracy against vLLM decodes. This release allows local users to run Inkling 2 models directly on CPU and GPU hardware.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1Thinking Machines releases Inkling model
Fireship · July 20, 2026 · Related
- 2Llama.cpp Adds Vision Support with MiniMax-M3
llama.cpp Releases · July 27, 2026 · Related
- 3vLLM v0.26.0 Release Enhances Model Support
vLLM Releases · July 27, 2026 · Related
- 4Meta unveils Muse Glimmer: a 30B parameter multimodal model
Hugging Face Blog · August 10, 2026 · Related
- 5Meta's Muse Glimmer Model Now Available
Ollama Blog · August 10, 2026 · Background
- 6v0.28.0rc2: DFlash2 Local Convolution Update
vLLM Releases · August 22, 2026 · Background
- 7llama.cpp adds vision target support for DFlash/Dspark
llama.cpp Releases · September 24, 2026 · Related
- 8Liquid AI releases DSpark drafter for vision models
Hugging Face Blog · September 24, 2026 · Related
More from llama.cpp Releases
Coding Toolscoding
llama.cpp b12398 optimizes Vulkan Flash Attention for AMD
This release targets the persistent performance gap on AMD hardware by optimizing Vulkan Flash Attention specifically for RDNA3 and RDNA4 architectures. The changes include reducing shared memory usage through buffer reuse and fixing critical stability issues like NVIDIA hangs and sparse FA underflow bugs. While there are no new model architectures, these low-level shader optimizations directly translate to faster inference speeds for local users on modern AMD GPUs. It is a quiet but necessary engineering update that tightens the ecosystem for non-NVIDIA accelerators.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12399 adds ROCm 10 and CUDA 13 builds
This release quietly extends llama.cpp’s hardware support to the latest NVIDIA and AMD stacks. By shipping binaries for CUDA 13.5 and ROCm 10.0, it ensures users on cutting-edge GPUs don't have to compile from source to get performance gains. The Vulkan build also fixes header inclusion issues that previously broke compilation on some Linux distributions. It’s a maintenance update rather than a feature leap, but it keeps the local inference runtime compatible with the fastest consumer and datacenter hardware available today.
llama.cpp Releases
Coding Toolscoding
llama.cpp b12400 release with CUDA 13 and ROCm 10
This is a routine maintenance release for llama.cpp that quietly extends hardware support to the latest NVIDIA and AMD stacks. By adding CUDA 13.5 and ROCm 10.0 binaries, it ensures local inference remains viable on cutting-edge GPUs without waiting for community forks. The build also includes OpenVINO and SYCL variants, reinforcing its role as the universal runtime for diverse accelerators. While no new model architectures are introduced, this update keeps the toolchain compatible with modern driver ecosystems.
llama.cpp Releases
More in Models & Labs
© GitHub ChangelogModels & Labscoding
Gemini 4 Hydrogen arrives in GitHub Copilot
Google’s latest Flash model is now generally available in GitHub Copilot, targeting agentic coding and multistep workflows. It offers improved instruction following over Gemini 3.8 Flash and is billed at provider list pricing under usage-based billing. The rollout is gradual across Visual Studio Code, JetBrains IDEs, and other platforms. Administrators can manage access through model policies in Copilot settings.
GitHub Changelog
© TechCrunch AIModels & Labsproductivity
ElevenLabs launches ElevenAI for teams
ElevenLabs is moving beyond voice synthesis into the crowded meeting intelligence space with ElevenAI. The app merges transcription, dictation across 90+ languages, and agent-driven follow-ups like drafting emails directly from conversation context. This positions their $22B valuation engine as a central knowledge hub for enterprises, aiming to replace standalone notetakers. It’s a strategic play to lock teams into their ecosystem using the recently launched v3 Scribe models.
TechCrunch AI
© TechCrunch AIModels & Labsmodels
Google deploys Gemini 4 Hydrogen for faster Search
Google is quietly shifting its search infrastructure by replacing older fast models with Gemini 4 Hydrogen, a tier designed specifically for speed and cost-efficiency rather than raw reasoning power. This model completes tasks 2.5 times faster than its predecessor while beating the more powerful Argon on specific intelligence benchmarks, proving that specialized efficiency can outperform brute force in high-volume workflows. By integrating this into Search's fan-out technique, Google now breaks down complex queries into subqueries with greater precision and speed, delivering richer visual responses without the latency of heavier models. This marks a strategic pivot toward agentic execution layers within consumer search, making AI Mode feel snappier for everyday multi-part questions.
TechCrunch AI