The Archive

Here's my article with **38 quant pairs** thoroughly benchmarked in KLD with **3 different Qwen 3.6 27B configs**: Q5\_K\_S + 64k context, IQ4\_XS + 64k context, IQ4\_XS + 128k context. This allows us to track not only how cache quantizations affects the precision in a vacuum, but also how it interacts with noise from the model itself. All benchmarks were done using my [BeeLlama.cpp](https://github.com/Anbeeld/beellama.cpp) fork, allowing to include a number of quant types that are not present in mainline llama.cpp: vanilla TurboQuant, TCQ 3-bit/2-bit, and q6\_0. [https://anbeeld.com/articl...

[Cuda 13.3 Downloads](https://developer.nvidia.com/cuda-downloads) [Release Notes](https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html) Anybody already tried llama.cpp with 13.3?

Here's the PR by pedapudi. https://github.com/ggml-org/llama.cpp/pull/21344 It's merge request has been denied so it will not be in mainline llama.cpp. The changes are so small that I just put them into whatever the current release of llama.cpp is. Read the PR for more info. It will only work with MOEs. Also, it gives the most boost at low context. As the context rises, the gain diminishes. Pedapudi explains why that happens in the PR. Here are some numbers. It really works well. The tiny amount of time it takes me to apply the code to the current release of llama.cpp is time well spent. ...

I've fine-tuned Qwen 3.5 0.8B on the dataset provided by Pangram with their EditLens paper. It's available via a [Chrome extension](https://chromewebstore.google.com/detail/slop-hammer/gfjdmhfokmhedlgfggmmgchpppmhkdgg); you can just click selected text and it's going to give you the probability distribution of how likely it is AI-generated. It takes under 1s on my M1 MacBook Pro. Pangram did release Llama 3.2 3B trained on their dataset, but I found this model slightly too legacy (too big for the capabilities). Qwen 0.8B (base) ended up being as good after roughly 20h of fine-tuning on a sin...

llama.cpp adds fast Walsh-Hadamard transform (FWHT) for CUDA, yielding 1–2% prompt-processing and 7–9% token-generation speedups with quantized KV-cache.

Financial Times reports Heretic tool removes guardrails from Meta's Llama 3.3 in <10 minutes; 3,500+ decensored variants downloaded 13M times.

llama.cpp PR addresses checkpoint creation inefficiency when context optimization tools modify conversation history in agentic workflows.

Community discussion on local LLM frontends and limitations of llama-server as default interface.

Benchmark of vision LLMs vs. OCR pipelines on 30 long, image-heavy PDFs from MMLongBench-Doc shows LlamaCloud + Azure premium achieving 59.6%–58.5% accuracy; agentic RAG and native PDF vision approaches compared on cost and accuracy.

llama.cpp server adds native tool support (shell execution, file ops) via experimental --tools flag.

Wanted to share a workflow I tested on a real flight, in case anyone else is trying to set up offline Claude Code. The core idea: using ollama to pull the needed model of what you need, and then use it to run claude code The setup, in order: 1. Pull a model on home wifi the night before. \`ollama pull &lt;model&gt;\` — \~9 GB for a 14B, \~17 GB for a 26B. Don't try this at the gate. 2. In Claude Code, point at Ollama. The cleanest path I found is wrapping it in two aliases: alias claude-local='ollama launch claude --model gemma4:26b' alias claude-cloud='claude' 3. Verify on the ground with...

User reports Apex quantization for Gemma 4 26B achieves 38 tokens/sec at 90K context on RX 9060 XT with llama.cpp.

User configures dual AMD RDNA GPUs (48GB VRAM) with llama-cpp via Vulkan for local inference.

BeeLlama v0.2.0 achieves 4-5x token throughput gains on RTX 3090 via DFlash optimizations for Qwen 27B and Gemma 31B models.

Community fork of llama.cpp optimizes MoE inference on 12GB VRAM by loading only active experts rather than full layers.

cHunter789 releases Qwen-27B IQ4_KS quantization (14.1GB) optimized for 16GB NVIDIA GPUs via ik_llama.cpp.

SupraLabs released Supra-50M, a 50M-parameter Llama-style language model trained on 20B educational tokens with competitive benchmark performance.

llama.cpp b9274 fixes VRAM leak in speculative decoding by properly freeing draft context and decoder resources on server sleep.

Qwen 3.7 open-weights model released; community discussion on LocalLLaMA highlights adoption momentum.

llama.cpp PR #22929 fixes prompt processing performance issue affecting OpenCode and Pi model inference.

User reports 110 tok/s inference speed on Qwen 35B with 12GB VRAM using ik_llama.cpp, outperforming standard llama.cpp after MTP merge.

User shares llama.cpp configuration for running Qwen 3.6 27B locally with ROCm acceleration and optimizations.

llama.cpp PR #23287 optimizes MTP (multi-token prediction) draft sampling by moving logic to backend, improving inference performance.

Early-stage Gemma 4 MTP compilation work-in-progress shared on LocalLLaMA.

LM Studio 0.4.14 adds MTP Speculative Decoding support via llama.cpp 2.15.0 for faster inference.

llama.cpp PR #23269 introduces MTP (Multi-Token Prediction) improvements for faster local LLM inference.

llama.cpp MTP speculative decoding merged; Qwen 27B/35B inference speedups 1.24–2.44× on consumer GPUs/APUs with quantization.

Just wanted to share that I'm pretty happy about Qwen 35b a3b agentic coding performance. I'm running the model in q80 quant, kv cache both q8\_0 as well, with 262144 in 4090 + 5060 ti, via llama.cpp backend with claude code pointing to localhost. For demo/data analytics purposes, it works pretty well. I haven't used it for large codebases, but it definitely is better than gemma4 26b in my use case. One thing that surprises me is that it seems to get better outcome in agentic coding, than chat. When using it with just chat UI, i found the code qwen35b provide a bit too clunky. I wonder o...

Llama.cpp performance improved 1.5–1.8x in recent updates; MTP and prompt processing issues partially resolved.

MTP KV cache quantization in llama.cpp Qwen models reduces VRAM overhead without apparent inference degradation.