Articles by tag: llama-cpp

Claude Code Leak: Source Code and Releases of Gemma 4 Qwen

Analysis of the 512K lines Claude Code leak, harness architecture, Gemma 4 and Qwen 3.6-Plus releases. Benchmarks, migrations, local launch. For middle/senior dev — read the details and build your own agents.

Running MoE GPT-OSS-120B on 6 GB GPU up to 30 t/s

Accelerating GPT-OSS-120B on a weak GPU with -cmoe and --fit. Efficient RAM usage for MoE models. Instructions for developers.

llama.cpp: ASR in Gemma 4 and Qwen3

llama.cpp integrated speech recognition Gemma 4 (E4B/E2B) and Qwen3. Launch on GPU, prompts, parameters for stability. Instructions for middle/senior dev.

File access in local AI agent: tools

Implementation of file access for local AI agent: permission gate, auto-verify, fixing KV-cache bugs. Technical details for developers. Study the architecture.

ruGPT3XL with 8k context: sparse attention

Expansion of ruGPT3XL to 8k tokens via sparse attention and fine-tuning. PPL 13.00 on gazeta, Triton/SDPA optimization ×1.85. Instructions for developers.

Restoration of ruGPT-3 XL for transformers

Learn how to convert legacy ruGPT-3 XL to HF format for fine-tuning and inference. Weight structure, custom classes, MERA benchmarks. Instructions for developers.