Tech

M1 Max ComfyUI test of WAI-illustrious-HSWQ v17.0 base vs WAI v17. Fast mode took 28-34s, quality 64s, with Kana LoRA and cafe background checks.

Tested ZONOS2 on an 8GB RTX 4060 Laptop (WSL2): the 15.3GB bf16 weights run via Windows system-memory fallback, a KV-cache override, and the CUDA toolkit at ~1/20 realtime. Plus a Japanese name-accent gotcha with A/B audio.

AMIXA Async's MIX ANIMA tested on M1 Max ComfyUI against anima-base-v1.0/JANIMA with Anima LoRA, two-character prompts, Turbo, and native cafe scenes.

Anthropic disabled Fable 5 and Mythos 5 for all customers on June 12, 2026. US export controls, foreign-national employees, AI Safety gaps, Microsoft limits, and the shaky Opus 4.8 fallback.

Tested on M1 Max ComfyUI: newly free JANIMA vs Hexer Minimal Toon Anima V1 vs anima-base, one character LoRA, same seed. Hexer keeps the outfit; JANIMA adds clothes but draws the quietest backgrounds.

Codex printed 'Selected model is at capacity. Please try a different model.' mid-task. It's model-side serving capacity, not context length, and the same thread resumed after a continue prompt — with the OpenAI maintainer's explanation (issue #17014) and which related issues stay open.

GitHub disabled 73 Microsoft repos after an Azure/durabletask commit. Miasma used Claude Code, Gemini CLI, Cursor, and VS Code config, not npm install.

LiteLLM 1.74.2-1.83.6 command execution via MCP stdio test endpoints is in CISA KEV. Patch to 1.83.7+ and Starlette 1.0.1+; BadHost can remove auth.

Claude Code CLI vs Codex CLI on a 7-test static-blog fixture: runtimes, estimated Claude Code cost, and the semantic diff Codex used to pass.

Laxhar's U1 trainer needs 32GB+ GPU, bf16 only — 4bit broke gen tower. Prefix offload keeps ~20GB peak. 8-step LoRA stack, A3B MoE compat, official training code gap.

AFM 3 splits into 20B on-device sparse (NAND-to-DRAM weight loading) and Cloud Pro on Google Cloud NVIDIA GPU. Three Google contexts, Foundation Models API opening, and what's still unreleased.

Tested LFM2.5-1.2B-JP-202606 on M1 Max 64GB. llama.cpp Q4_K_M: 208 tok/s decode, JSON intact, model name hallucinated (LFM→FDM). Q8_0: 157 tok/s, no hallucination. Tool calls broken via GGUF.