llama.cpp merges native Blackwell NVFP4 support

// 45d agoOPENSOURCE RELEASE

llama.cpp merges native Blackwell NVFP4 support

llama.cpp has merged preliminary SM120 native NVFP4 MMQ support, bringing hardware-native FP4 inference to Blackwell-class GPUs. The post also notes that GGUF builds are already appearing for models like Gemma 4, Nemotron Cascade 2, and Qwen3.5 in NVFP4 form.

// ANALYSIS

This is a meaningful infrastructure step, not just another quantization tweak: llama.cpp is moving from "can load the format" toward actually exploiting Blackwell silicon the way NVIDIA intended. It’s still preliminary, but it should matter immediately to anyone chasing better throughput-per-watt on local or semi-local rigs.

–The merge targets SM120 Blackwell GPUs, so the win is tied to newer NVIDIA hardware rather than a broad across-the-board speedup
–Native NVFP4 support lowers the gap between model packaging and kernel support, which is why GGUF variants are already surfacing so quickly
–For local inference users, this strengthens llama.cpp’s position as the first stop for bleeding-edge quant formats and vendor-specific hardware features
–The "preliminary" label matters: expect rough edges, model-by-model quirks, and a period of rapid follow-up fixes
–This is especially relevant for MoE and larger models where memory bandwidth and quantized math are the bottlenecks

// TAGS

llama-cppgpuinferenceopen-sourcellmself-hosted

DISCOVERED

45d ago

2026-04-29

PUBLISHED

45d ago

2026-04-29

RELEVANCE

9/ 10

AUTHOR

ggonavyy

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL15m ago

Moonshot AI has officially released Kimi K2.7-Code, an open-weights coding model optimized for long-horizon software engineering and cost-efficient agentic reasoning.

Moonshot AI has officially released Kimi K2.7-Code, an open-weights Mixture-of-Experts coding model featuring 1 trillion parameters and a 256K context window. Optimized for long-horizon software engineering tasks like codebase-wide refactoring and debugging, the model achieves a 30% reduction in reasoning-token usage compared to its predecessor. Kimi K2.7-Code supports multimodal inputs, runs in a dedicated reasoning-heavy thinking mode, and is available for developers via Hugging Face, Ollama, and the Kimi API.

TUTORIAL47m ago

Seedance 2.0 workflow animates consistent characters

AI creator Aimi Kōda shared a step-by-step generative AI workflow titled "Surf on the Clouds" that coordinates Midjourney, GPT Image 2, and Seedance 2.0. The tutorial explains how to generate a stylized character in Midjourney, build a structured 16:9 character identity sheet using GPT Image 2, and animate the assets using Seedance 2.0 to maintain visual and narrative consistency across scenes.

MODEL1h ago

Claude Fable 5 overshadows Claude Opus 4.8

The rapid succession of Anthropic's model releases has left Claude Opus 4.8—which debuted just two weeks ago as a major frontier model—largely forgotten in the wake of Claude Fable 5. Fable 5's introduction as the first generally available 'Mythos-class' model has generated massive hype due to its superior score of 80.3% on SWE-bench Pro and impressive multi-step autonomous planning, completely shifting the AI community's focus and discussions away from the incremental updates of Opus 4.8.

llama.cpp merges native Blackwell NVFP4 support