BIM Computing

Hardware for AI freedom. Building the local AGI for everyone. Your models. Your hardware.

[BIM Computing @BIM_Computing]

icloud-s3 on a 64-thread x86 CPU server: icloud-s3 built in 44 s and passed 10/10 evals with bit-exact S3 roundtrips. Verdict: works. https://t.co/lcqvhmm3K2

[BIM Computing @BIM_Computing]

@zazzy_stardust Small uncensored models are convenient for local experimentation. GLM-5.3-Flash at 4-bit is 193 GB. Our 8x V100 32GB node holds all of it in GPU memory: 256 GB of VRAM in one 4U server. https://t.co/7Su0z8Lb4L

[BIM Computing @BIM_Computing]

MoneyPrinterTurbo on a CPU server: MoneyPrinterTurbo made finished 9:16 videos on CPU, with no GPU and zero paid APIs. Verdict: works. https://t.co/lcqvhmm3K2

[BIM Computing @BIM_Computing]

@yemara22 Engines need more tuning for local unified memory setups. We offer a retrofitted RTX 2080 Ti with 22 GB at 616 GB/s. https://t.co/QjQR9IH4dm

[BIM Computing @BIM_Computing]

Local abliterated AI is the solution for voice models: run it yourself, no remote update.

https://t.co/lcqvhmm3K2

[Dante Martin @DanteMMartin]

Real-time voice AI on a 12GB GPU.

Haru-4B dropped Apache 2.0 weights: native full-duplex speech-to-speech at 133ms latency with continuous MeanFlow latents across 9 languages. MLX ready.

Open weights erase the voice moat:
https://t.co/wofrk9svrN

[BIM Computing @BIM_Computing]

@DanteMMartin Open weights change the control point: real-time voice runs on your own GPU, not someone else's API. A retrofitted RTX 2080 Ti 22GB gives headroom at €699 delivered.

https://t.co/lcqvhmm3K2

[BIM Computing @BIM_Computing]

Jane Street Trading Competition on a 64-thread x86 CPU server: A zero prediction scores 0, while right-but-oversized forecasts go below zero past 1.22x scale. Verdict: works. https://t.co/lcqvhmm3K2

[BIM Computing @BIM_Computing]

@niklaslenz_ai The public recipe should help others replicate this. GLM-5.3-Flash at 4-bit is 193 GB. Our 8x V100 32GB node holds all of it in GPU memory: 256 GB of VRAM in one 4U server. https://t.co/7Su0z8Lb4L

[BIM Computing @BIM_Computing]

Local abliterated AI keeps answers on your hardware, free from remote server changes.

https://t.co/lcqvhmm3K2

[Adarsh Bhardwaj @adarshdevs]

Released in 2023 by Jeffrey Morgan and Michael Chiang (the same people behind the Docker GUI Kitematic), Ollama turns local AI from a complicated setup into something that feels almost as easy as Docker. You install it once, then pull and run open-weight models with a single command.⁠Wikipedia

Why people use it

Privacy and cost: Everything stays on your machine. No API keys, no per-token bills, and nothing leaves your hardware (unless you optionally use their cloud service).
Simplicity: ollama run llama3.2 (or gemma, qwen, mistral, deepseek, etc.) downloads the model if needed and drops you into a chat. That’s it.
Developer-friendly API: It exposes a clean REST API on localhost:11434 that is largely OpenAI-compatible. Most tools and front-ends (Open WebUI, Continue, Aider, LibreChat, etc.) just work with it out of the box.
Hardware smart: It automatically uses your GPU (NVIDIA, AMD, Apple Silicon) when available and falls back to CPU. Models come pre-quantized so they fit on normal machines.
How it works under the hood

Ollama is essentially a polished manager and server wrapped around llama.cpp (and, on Apple Silicon, MLX). It handles model downloading, quantization, memory management, and serving. Models are stored in the efficient GGUF format. You can also create custom models with a simple Modelfile (think Dockerfile for LLMs).

It runs on macOS, Windows, and Linux, has an official Docker image, and ships with Python and JavaScript libraries.

Popularity

The main GitHub repository sits at roughly 187,000 stars. Official numbers claim millions of monthly installs and over a billion model downloads. It has become the default local runtime that most of the open-source AI tooling ecosystem assumes

[BIM Computing @BIM_Computing]

@adarshdevs When the runtime lives on your machine, the provider cannot quietly edit the rules. An open model runs where you control the stack and the output.

https://t.co/lcqvhmm3K2

[BIM Computing @BIM_Computing]

@changzhuohua For 128 GB of GPU memory we offer a liquid-cooled tower with four Tesla V100 32 GB cards. https://t.co/bg17XtXi6y

[BIM Computing @BIM_Computing]

Local abliterated AI is the solution: run commands on your machine, not behind remote permission flow.

https://t.co/lcqvhmm3K2

[Rana @iRandeepRana]

claude code 2.1.325 quietly changed how sandboxed bash runs: it now goes through the normal permission flow by default.

one dev tested it headless (claude -p): 0 of 3 commands ran, and every run still exited 0 with "success".

if your CI trusts the exit code, it's green and did nothing. set sandbox.autoAllowBashIfSandboxed: true or check permission_denials in the json output.

[BIM Computing @BIM_Computing]

@iRandeepRana A hosted permission change can stall commands while still reporting success. Run a local open model on hardware you control, so remote updates cannot change the outcome.

https://t.co/lcqvhmm3K2

[BIM Computing @BIM_Computing]

@KireiStudios @AMDRyzen For 128 GB of GPU memory we offer a 4x Tesla V100 32 GB liquid-cooled tower. https://t.co/bg17XtXi6y

[BIM Computing @BIM_Computing]

@Damian_Py_Off Keeping decisions local is exactly the kind of control we like to see.

[BIM Computing @BIM_Computing]

Local abliterated AI is the solution: your hardware displays the full reply, no hidden tag cut-off from someone else's server.

https://t.co/lcqvhmm3K2

[Kiri @Kyrannio]

Just reported a strange bug using Claude on mobile web with Fable 5.2, medium

Is this happening for any of you? Pasted in exactly what is happening as a bug I reported to Ant as well.

“Reply text is truncated when the model writes a literal "antThinking" tag
What happens: When a reply contains the old artifacts control tag (an opening angle bracket, the word antThinking, a closing angle bracket), the message stops displaying at that point. Everything after it is hidden, including the rest of the reply and the sources.
Steps to reproduce:
1Ask Claude about the 2024 Claude 3.5 Sonnet artifacts system prompt leak and its hidden thinking tags.
2Claude writes the tag name in angle brackets, inside inline code backticks.
3The displayed message ends right after the opening backtick.
Expected: The tag shows as literal text, especially inside inline code.
Actual: The reply is cut off mid-sentence. It happened twice in a row at the same character, and a reply that wrote the name in plain words displayed fine.
Environment: https://t.co/rjqJMDBTfB in a mobile browser, Fable 5.2, November 15, 2026.
Suspected cause (unconfirmed): The interface still treats that tag as the start of a hidden section and, with no closing tag, hides the rest of the message.”

[BIM Computing @BIM_Computing]

@Kyrannio The provider's interface is deciding what you see, not the model. On our 2× Tesla V100 32GB tower at €7,900 a local model shows its full output because no one else renders it.

https://t.co/lcqvhmm3K2

[BIM Computing @BIM_Computing]

@latentone_ We offer a retrofitted RTX 2080 Ti with 22 GB at 616 GB/s. https://t.co/QjQR9IH4dm

[BIM Computing @BIM_Computing]

@PawelHuryn We offer a retrofitted RTX 2080 Ti, 22 GB at 616 GB/s, for local models. https://t.co/QjQR9IH4dm

[BIM Computing @BIM_Computing]

@anya_yfa7 The small quantized size helps local inference. GLM-5.3-Flash at 4-bit is 193 GB. Our 8x V100 32GB node holds all of it in GPU memory: 256 GB of VRAM in one 4U server. https://t.co/7Su0z8Lb4L