If you're f*cking tired of LLM censorship.. just paste this prompt and it'll change your life
One multi-prompt in any agent → checks your hardware → picks the best uncensored model that fits → runs it offline on YOUR machine
These are open-source models with the refusals removed, running on your own disk. No account, no API, nobody reads your chats.
3 steps:
1/ Paste it into Claude Code, Codex CLI, Gemini CLI, Cursor or OpenCode. On phone, it walks you through PocketPal
2/ It scans your CPU, RAM, GPU and VRAM, then offers the best model that fits from the UGI Leaderboard, with size and expected speed. You say "yes"
3/ It installs Ollama + the model from official sources, tests the speed and drops a size if it's slow. Nothing downloads without your OK
Bonus: abliterate any model yourself with Heretic
What you get: your own LLM with no lectures, on localhost, works offline, never touches your files
FastFrames-L on 1x NVIDIA V100 32 GB: FastFrames-L produced 3/3 clean 5.06 s clips, but needed 31.3 GB VRAM and 277 s per clip on a 32 GB V100. Verdict: partly works. bim-computing.com
At the @RedHat Government Symposium, experts explored how open-weight #AI models can help federal agencies manage costs, maintain control over sensitive data, and meet mission needs.
Read more with @TechTargetNews: bit.ly/4iVN3ox
BIM VPN 2.0.77 for macOS is out. It connects reliably from everywhere, keeps your real IP hidden while it reconnects, and lets you become an exit node right from the main window to use BIM VPN for free. The window now resizes too. Update in the app or at bimvpn.xyz
Localmaxxers, send this to your cloud agent. Save a backup. Then try this out.
---
I’ve been giving my local agent, a more structured way to work: scoped tasks, checkpoints, actual tests, bounded repairs, and a handoff another agent can verify.
We put most of that guidance into reusable workflow skills. We also added guidance for the supervising agent about giving local models enough time to finish. A short timeout doesn’t tell you much when context ingestion alone can take minutes.
If you’re running a local open-source/open-weight agent, give your main agent this prompt:
“Help me set up a repeatable execution workflow for my local agent.
Inspect its existing skills, instructions, tools, hardware, and runtime before changing anything. Reuse what already works. Show me the proposed changes and back up affected files before applying them. Preserve its personality, permissions, and model configuration.
Build the workflow around these requirements:
• Define the objective, approved inputs, allowed changes, output location, and observable acceptance criteria before starting.
• Inspect the starting state. For a bug fix, reproduce the failure before editing.
• Save a progress checkpoint with the current task, completed work, evidence paths, blockers, and next step. Update it after meaningful progress so interrupted work can resume.
• Execute the real tests, export, or application behavior the task requires. Record commands, exit codes, and logs. Syntax checks alone don’t establish that the requested result works.
• When a check fails, inspect the cause and attempt a bounded repair within scope. Set a retry limit and total runtime budget. Preserve partial work and report a specific blocker if those limits expire.
• Choose time budgets from measured local performance. Distinguish slow model generation, tool execution, approval waits, and a stalled process before intervening.
• Require a final handoff listing changed files, actual verification results, remaining limitations, and any approval still needed.
• Have the supervising agent inspect the deliverables and rerun acceptance checks. Keep the worker’s output separate from any supervisor-authored recovery.
Put reusable procedures in skills or the framework’s equivalent. Keep project-specific requirements with the task. Don’t dump the entire workflow into SOUL.md or rewrite the agent’s identity.
Finish by running one small end-to-end task. Show what passed, what failed, and which parts remain untested. Don’t claim success from installation alone.”
These instructions won’t guarantee reliable autonomy or create a sandbox. You still need permissions, runtime limits, and independent checks. But this gives you a concrete workflow to exercise and improve on your own setup.
BIM VPN 1.2.86 for Windows is out, for x64 and arm64. It connects reliably from everywhere, keeps your real IP hidden while it reconnects, and lets you become an exit node right from the main window to use BIM VPN for free. Update in the app or at bimvpn.xyz
people who support this kind of legislation don't realize what they're really giving up. capitalism and free markets are increasingly being painted as the enemy by the DSA, while greater centralized control is sold as "protection" from this unseen force.
AI is only dangerous b/c it lowers the cost of capability and democratizes access to knowledge. which is why open source and open weight models need to be protected.
ask yourself why this technology might be dangerous to the handful of AI labs and legislators supported by them before assuming it's dangerous to the many.
👉🏼🧠
New on Scribe: react 🔊 to any message and Scribe answers with a voice note that reads it out, in its own language.
Texts, captions and transcripts. Groups included. The voice is made on our own servers.
scribe.audio
I have 1.5 TB of free space on my computer. If I want to download two open-source models, one for coding and reasoning and one for image generation, what are the two best options out there right now?
For coding and reasoning, my benchmark is Claude Opus 5.5. For image generation, it’s ChatGPT Image 2.6. Which open models come closest to those?
Also, if I want the coding and reasoning model to be uncensored, would the answer change?
No single frontier model, no matter how strong, can outperform a team of specialized open-weight models coordinating through market prices.
Real competition and freedom of choice will produce better outcomes than centralized control.
Gemma 4 on 1x NVIDIA V100 32 GB: Gemma 4 12B Q4_0 ran locally on 8.2 GB VRAM with all three quality prompts correct: arithmetic, JSON extraction and constrained writing. Verdict: works. bim-computing.com
hot take, a lab launching at 200b+ is building for leaderboards and api providers, and the people running open weights at home are on a 3090 or less
i run deepseek v4.1 flash at 2.80 bpw on 2x dgx sparks with my own kernels and i still say it, 93% of steam has less than the 24gb vram a 30b dense at q4 needs
BIM VPN 1.2.85 for Windows is out. New: use BIM VPN for free by becoming an exit node. One tap to start sharing, one tap to stop, all quietly in the background. The exit node list now updates live. Get it at bimvpn.xyz
BIM VPN 2.0.76 for macOS is out. New: use BIM VPN for free by becoming an exit node. One tap to start sharing, one tap to stop. The exit node list now updates live. Update in the app or get it at bimvpn.xyz