LuffyTheFox's Qwen3.6 Merges Uncensored AI and Tool-Calling Into One Local Model
A community-built GGUF release fuses an uncensored Qwen3.6-35B MoE base with Hermes function-calling weights and a custom tensor-denoising pass.
20 stories
A community-built GGUF release fuses an uncensored Qwen3.6-35B MoE base with Hermes function-calling weights and a custom tensor-denoising pass.

A community-released GGUF blends an uncensored Qwen3.6 MoE base with Hermes function calling, using a math-based tensor repair method called Genesis.

A community-built GGUF stacks an uncensored Qwen 35B MoE with a Hermes agent transfer and a tensor-level noise repair pass called Genesis.

ByteDance open-sources SandboxFusion, a self-hosted code execution sandbox that runs 20+ languages and powers both LLM evaluation and RLVR training pipelines

Prime Intellect overhauls its RL environment stack with composable tasksets, harnesses, and a smarter trace format that makes long-horizon agentic training actually feasible.

OpenAI's Agents SDK gains Sandbox Agents in v0.14 -- giving AI agents a real filesystem, shell access, and resumable workspace to tackle long-horizon tasks

China Mobile's JT-4.1 Flash 236B A21B scores 39 on the AA Intelligence Index, closing the gap to China's AI frontier with a massive leap from its predecessor.

Cursor 3.11 adds side chats for parallel exploration, full-text transcript search, and new cloud agent hooks for self-correcting loops

Epoch AI analyzed OpenAI's Codex repo and found 8% of engineer workdays in Q2 2026 exceeded 24 hours of estimated human effort, up from 2% a year ago.

Google's Willow chip runs Quantum Echoes 13,000x faster than supercomputers, producing the first verifiable quantum advantage with real molecular applications

Perplexity swaps in xAI's freshly-launched Grok 4.5 as the orchestrator brain of Computer, beating every rival configuration on WANDR at roughly half the cost of Claude Opus 4.8

OpenAI doubles its biosafety jailbreak reward to $50K and turns a one-time bounty into a permanent program, as GPT-5.6 earns a 'High' bio-risk rating across all three model tiers for the first time.

GitHub's Copilot code review got worse after a tool upgrade — until they rewrote the instructions, cutting costs 20% with no quality loss

Upstage's Solar 31B now runs at up to 2,000 tokens/sec on Cerebras hardware, making deep research queries across hundreds of sources feel nearly instant.

humans& open-sources a stable NVFP4 reinforcement learning recipe that matches BF16 training dynamics while delivering major throughput gains on Blackwell GPUs

Anthropic's claude-quickstarts repo packs six runnable agent templates — from browser automation to autonomous multi-session coding — all free under MIT.

Meta's new kernel fusion techniques hide up to 90% of LayerNorm/RMSNorm latency inside GEMM and Attention kernels, cutting full Attention block latency by 35%

Google DeepMind and Google Labs add Street View grounding to Project Genie, letting users build interactive AI worlds anchored in real locations
Perplexity Computer now shows credit spend broken down by model, available to all consumer and enterprise users in Account Settings.

Google AI Studio now lets you give deployed apps a clean atelier.ai.studio URL instead of a random Cloud Run slug