DeepSeek-V4.1-Flash at 27 Tokens/s on One RTX 4090: Our Xolotl Engine
Our open-source engine runs all of DeepSeek-V4.1-Flash on one RTX 4090 + 90 GB RAM at about 27 t/s with 512K context, plus GLM-5.3-Flash and Qwen3.8, behind one OpenAI/Anthropic API for coding agents. Where the 15,360 experts live, where each token's 38 ms goes, how agent turns got fast, and the per-expert quants that feed it.
Read Article →Strata-GLM: A 320B Model at 25 t/s on One RTX 4090 — From 1.75 t/s in Three Days
We ported GLM-5.3-Flash to our own Strata engine on a 24 GB card + 90 GB RAM: 25 t/s decode, ~400 t/s prefill, 512K context, uncensored with a runtime abliteration LoRA. Flat routing, an abliteration loop bug, a crash at 19K tokens, and the profile that proved our own theory wrong.
Read Article →Flash-Next vs a 27B on One RTX 4090 (24 GB): What 30 Hidden-Graded Agent Runs Show
A 177B MoE and a dense 27B on a single 24 GB card, through a real agent harness with hidden graders: vision, browser, desktop, ComfyUI and from-scratch interpreters. The 27B scored 64/64 at ~150 t/s; Flash-Next was slower but the only one to ace every hidden edge case.
Read Article →Strata: Flash-Next at 76 t/s on a Single RTX 4090 (24 GB), Then Unrefusing It With a Direction
A purpose-built engine decodes Qwen3.8-Flash-Next at 76 t/s on one 4090 — 2× llama.cpp. Then a single directional control vector cuts its refusal rate from 100% to 27% with throughput intact and zero weight edits.
Read Article →Runtime Abliteration: Removing Refusal Without Touching a Weight
Our first control vector failed — a "refusal direction" built from the wrong contrast set just shuffles which topics get refused while wrecking coherence. Rebuilt from real refusal pairs, the same code at the same scale took a 12-prompt boundary probe from 2 refusals to zero, with thinking intact and no measurable capability loss. Plus the truncation trap that made a broken config score as perfect.
Read Article →Qwen3.8-Flash-Next on One RTX 4090: 20 to 65 t/s MoE Decode
An expert cache plus --load-mode none tripled decode speed for a 220B mixture-of-experts model on a single 24GB GPU. Plus the memory-mapping trap that cost us a day, n-gram speculative decoding that adds 50% on code for free, and a model file that had been quietly corrupt for weeks.
MiniMax H3: The New King of AI Video Generation
Text-to-video, image-to-video, and reference-to-video with consistency that finally holds up. How H3's structured prompting works, plus our open-source ComfyUI-H3O node pack that makes it practical.
Read Article →QuetzaCodetl: The Uncensored Claude Code — Run NSFW Coding AI Locally
A local-first fork of Claude Code with zero content restrictions. No API keys, no logging to Anthropic, no filters — abliterated models on your own GPU. Same tool-use architecture, none of the censorship.
Read Article →The Fable Distillation: Making a $3 Model Work Like a $30 Model
A behavioral overlay that brings frontier-model discipline to local and budget models — goal-driven work, self-verification, honest state reporting. No fine-tuning required. Available now as a downloadable prompt pack.
Read Article →Giving a Blind LLM Eyes: Desktop Control Without a Vision Model
Let your AI agent see your screen, click buttons, and automate any desktop app. No vision model needed. $29 AUD.
Read Article →I Built My Own Search Engine Because Google Stopped Being One
Odin: a self-hosted discovery engine that maps the internet's authoritative hubs and fetches content live. 250K+ pages across 10K+ domains, MCP server for AI agents, zero tracking, no censorship.
Read Article →Fused TBQ4 Flash Attention: 82 tok/s with Lossless 4-bit KV at 200K
We fused TurboQuant dequant directly into the flash attention kernel — reading raw TBQ4 blocks inline via centroid lookup in the FWHT-rotated domain. 82+ tok/s with lossless 4.25 bpv KV at 200K context on RTX 4090. Nobody else has done this.
Read Article →Turbo4 KV Cache: Better Than Q8 at Half the VRAM
Trellis-Coded Quantization benchmark: Turbo4 scores 100/100 vs Q8_0's 91 on hardened agentic benchmark. Lossless FP16 quality, 40 t/s, 256K context on RTX 4090 24GB.
Read Article →Q4 KV Cache: Surprisingly Viable
Benchmarks show Q4 KV cache producing faster, higher-quality code than Q8. The conventional wisdom about Q4 being unusable may be wrong.
Read Article →IQ2 vs IQ3 Quantization: 2x Speed, Same Quality
Comprehensive RTX 4090 benchmark: IQ2_XS hits 86 t/s vs IQ3_XXS at 44 t/s. Quality testing across coding, debugging, and agentic tasks shows negligible difference.
Read Article →Maxing Out Qwen3-Coder-Next Abliterated: 94 t/s
The abliterated version hits 94 t/s at 168K context with Q8 KV cache — over 5x faster than the base model. Complete optimization guide with red teaming verification.
Read Article →Qwen3-Coder-Next: IQ2 vs IQ3 Benchmarks
IQ2_XXS achieves 22 t/s on RTX 4090 at 200K context — 85% faster than IQ3 with no measurable quality loss. Full benchmark data and configuration.
Read Article →Running Uncensored AI Locally: My PRISM Setup
How I set up GLM-4.7-Flash locally with web search and vision capabilities. No content filters, no API subscription, just my hardware doing what I tell it to.
Read Article →Get AI Tips in Your Inbox
Subscribe for tutorials on local AI, Stable Diffusion, LoRA training, and Claude Code workflows.
More Articles Coming Soon
We're preparing additional tutorials and guides:
- Stable Diffusion Mastery — Advanced prompting and workflow optimization
- LoRA Training from Scratch — Create custom models for any subject
- LLM Fine-tuning Guide — Adapt open-source models for your domain
- AI Automation Pipelines — Build end-to-end workflows with n8n
- Claude Code Pro Tips — Advanced usage and MCP development
Have a topic you'd like us to cover? Let us know