QwenMix-3.7 Merges Qwen 3.8 and 3.6
QwenMix-3.7 is an experimental merge of the Qwen3.8-27B and Qwen3.6-27B models. The creator reports that it passed a basic smoke test, but no meaningful evaluation has been performed yet.
Search
Tag: #gguf28 results
Luth-2 Sets a New French SLM Benchmark
Luth-2-0.8B and Luth-2-2B are new non-reasoning French language models designed for local and on-device use. The developers report state-of-the-art results for their size across French benchmarks, with models, GGUF files, training data, and code available now.
NVIDIA Brings Its Speech Stack On Device
NVIDIA’s NeMo-Speech.cpp enables local, GGUF-quantized execution of speech recognition, text-to-speech, and codec components. The listed stack includes Magpie-TTS, Nemotron Speech, Nemotron ASR, Parakeet models, and NanoCodec.
Qwen3-TTS Voice Cloning Joins Mainline llama.cpp
Mainline llama.cpp now supports local Qwen3-TTS-12Hz-1.7B-Base voice cloning through the llama-tts binary. It accepts WAV or MP3 speaker references and supports 10 languages, but server integration and broader model support are still unfinished.
Cohere 30B A3B model GGUF release by Unsloth
Unsloth has released a GGUF version of the new Cohere 30B A3B model, making it accessible for local inference. This release is likely linked to recent updates in the llama.cpp ecosystem for improved model compatibility.
⚙️
Arc Pro B70 Trails RTX 3090 in llama.cpp Benchmarks
Benchmarks compare Nvidia RTX 3090 and Intel Arc Pro B70 using llama.cpp on Vulkan and SYCL backends across multiple GGUF models. Arc Pro B70 shows 71% slower prompt processing and 53% slower token generation on average versus RTX 3090. SYCL backend improves some Arc results over Vulkan.
Qwen3.6 27B GGUF Usable for Complex Coding
Qwen3.6-27B-UD-Q6_K_XL.gguf runs at 50 tok/s on RTX 5090 with 200k context via llama.cpp. Excels on difficult planning tasks, promising vs prior local models.
Qwen3.6 Autonomously Builds Tower Defense Game
A user tasked Qwen3.6-35B with building a tower defense game using screenshots from MCP, and it successfully implemented and tested features like upgrades. The model self-detected and fixed bugs in canvas rendering and wave completions. High excitement for the upcoming Qwen Coder model.
Gemma 4 31B SpecDec +29% Speedup
Speculative decoding with Gemma 4 E2B draft model boosts Gemma 4 31B inference by 29% on average, reaching 50% on code generation. Key issue was fixed GGUF metadata mismatch causing token translation overhead. Achieves high acceptance rates on structured tasks like math and code.
⚙️
ZINC: Zig LLM Inference for AMD GPUs
ZINC is a new LLM inference engine written in Zig, enabling 35B models on $550 AMD GPUs via Vulkan. It loads GGUF models, achieves 7.1 tok/s on RDNA4, and addresses AMD consumer GPU gaps. Repo at github.com/zolotukhin/zinc.