New quantization formats, speculative decoding and MoE offloading aim to run large models on single GPUs or CPUs, but runtime support often lags behind model releases.
ISTA-DASLab's release of GGUF quantized weights for DeepSeek-V4.1-Flash-1017 is not just a quantization milestone. It exposes the real bottleneck in local deployment: runtime integration. With llama.cpp still lacking official support and references p...
KLD is often treated as a universal metric for ranking quantized models, but it only measures the average divergence between the probability distributions of the base model and the quantized one. The number is fragile: it depends on the dataset, is e...
ISTA DASLab has released three GGUF builds of DeepSeek-V4.1-Flash-1017 with GSQ and RCO quantization, at 1.75 to 2.25 bits per weight over the 551.9 billion backbone parameters. A shared 110.6 GB Engram file stays memory-mapped on disk. The model run...
An experiment with a GGUF Q4 LLM in WebGL2 breaks on tokenization, not linear algebra. The advice to clone llama.cpp signals that in local and self-hosted deployments the differential lies not in the engine but in integration with data and processes....
ISTA-DASLab has published GGUF quants of DeepSeek-V4.1-Flash-1017 using the GSQ-RCO quantization approach. However, llama.cpp does not yet support the architecture: the relevant pull requests are still open. The model card mentions Strata, but the or...
A developer on Reddit describes an attempt to build a browser-based LLM inference engine using WebGL2 and GLSL, running a GGUF model in Q4 quantization. Tokenization becomes an endless debug hell. The advice: clone llama.cpp. AI-RADAR examines what t...
Kitten-TTS 2 brings 1.58-bit ternary quantization to a 1.7 billion parameter model that runs in real time on CPU. Self-hosting becomes realistic without GPUs, shifting TCO from VRAM to software execution quality. For organizations under data sovereig...
KittenML has released code and weights for Kitten-TTS 2, a TTS model with 1.58-bit ternary quantization and 1.7 billion parameters. It runs in real time on CPU, performs zero-shot voice cloning from under five seconds of audio, supports more than 20 ...
Specter combines quantization, speculative decoding, and expert offloading to speed up mixture-of-experts inference on a single GPU. A lightweight quantized draft model runs on-device and anticipates expert usage with high fidelity, reducing PCIe bot...
KRAFTON AI releases Raon-1 mini preview, a 30B MoE LLM with 4B active parameters, focused on Korean conversation and agentic workflows. With Apache 2.0, speculative decoding support, and NVFP4/MXFP8 quantization, inference can run on a single NVIDIA ...
The new EAGLE-4 method posts a 3.59× speedup over vanilla decoding in published benchmarks, ahead of DFlash2 (2.80×), DSpark (2.57×) and MTP (2.22×). Support is currently limited to Qwen3.8-27B, and integration with llama.cpp and vLLM is missing. For...
A user documents a jump from 9 to 48 tokens/s after switching to STRATA on CachyOS with ready NVIDIA drivers, running Qwen 3.8 Flash Next at 3 bits. The 8 GB GPU handles a 32k context and reaches 42k, staying 130 MB from crashing. The case highlights...
