BIM Computing (@BIM_Computing) header
BIM Computing (@BIM_Computing) profile picture

China GPU & sourcing / AI hardware: modded GPUs and AI server nodes. Part of BIM Groupe.

Shenzhen
Joined February 2024
Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)Photo by BIM Computing (@BIM_Computing)
THE BEAST IS ALIVE Benchmarking the BIM TerraNode 256: 8x NVIDIA V100 32GB (256 GB VRAM), 256 GB RAM, 4x 2 TB in RAID 10. Results will follow. Get yours now: bim-computing.com/products/8…
2
2
11
Local abliterated AI is the solution: your own hardware keeps the weights and removes the provider refusal list. bim-computing.com
10 UNCENSORED AI MODELS YOU CAN RUN WITH 16GB VRAM Qwen3.8 27B Abliterated huggingface.co/huihui-ai/Hui… Gemma 4 12B Uncensored huggingface.co/models?search… Ornith 1.5 9B Uncensored huggingface.co/models?search… Qwen3.5 9B Uncensored huggingface.co/localgstar/Qw… Dolphin Mistral 24B huggingface.co/models?search… LFM2.5 24B A2B Uncensored huggingface.co/OS-Software/L… Qwen3 VL 8B Abliterated huggingface.co/models?search… Mistral NeMo 12B Abliterated huggingface.co/models?search… GPT-OSS 20B Heretic huggingface.co/models?search…) Llama 3 8B Abliterated huggingface.co/models?search… use GGUF quantized versions to run locally check VRAM requirements before downloading bookmark for later use
2
10
49
Local abliterated AI is the solution: your own GPU, your own weights, no remote refusal list. bim-computing.com
All AI censorship is designed to prevent you from creating something popular without paying thousands of dollars for it. It's all about controlling who speaks and what they say. Clever people will manage to do and say what they want any way.
2
11
30
RTX 3080 unlocked, on the 35B MoE model we serve. Prompt tokens/s: 4.4k: 2,366 stock → 3,057 unlocked + Marlin FP4 kernel 32k: 2,457 → 5,273 with Marlin FP16 math, FP32 accumulate: 64 → 116 TFLOPS Same answers. Output at 4.4k: 50 → 41 tok/s. bim-computing.com
2
9
73
Local abliterated AI is the solution: run uncensored models on hardware you control, bypassing server-side restrictions. bim-computing.com
Anthropic decides what you're allowed to build now. On Oct 27th it pushed a new rule to Claude Code from its servers, with no update for you to install and no setting to turn it off. Anthropic now tells Claude not to recreate a specific existing game, even if you rename it. If Anthropic can do this to games, it can do it to whatever you're building.
2
9
49
AI companies now decide what you can think and do. In R&D or business, you can't keep proving to an AI your work is legitimate. Abliterated DeepSeek V4.1 Flash 1017: 45 on the Intelligence Index, close to Opus 5 (51). The future of AI freedom is local. bim-computing.com
10
151
Tried FreeToken from UC Berkeley on an old RTX 3080 10GB + DDR4. Qwen3.6-35B-A3B (MoE) runs at 68 tokens/s with the full 262K context. Same model in llama.cpp: 34 tok/s. A dense 27B offloaded to RAM: 2.9 tok/s. With MoE, RAM offload finally works.
3
1
11
697
That's probably why we got strange requests from banks with @BIM_Computing We do trading of modded GPUs from China, they asked us to perform KYC on our customers (wtf why need KYC to buy a graphic card?) and eventually @vividmoney_fr froze our customers' funds.
I was kidding about banning the 5090 wtf ahhhh
3
1
10
214
Vivid: a scam? @vividmoney_fr froze €47,635 of funds our customers paid us. Account closed "for cause" on 30/09, no reason given. Hold: "3 months" in their notice, "28 days" per support, release "not automatic". Formal notice sent. @vivid_en @dgccrf #fintech #neobank
12
8
2
18
440
Qwen3.8-27B Q4_K_M (15.6 GB weights) — decode tok/s vs price. 11 configurations: VRAM, memory bandwidth, four context/KV combinations, and whether 256k fits. Highlighted: Tesla V100 32GB, Tesla P100 16GB, RTX 2080 Ti 22GB.
1
1
1
63
Q8_0 (29.5 GB weights) — and Apple's unified-memory bandwidth per chip model. Plus why 32 GB is enough for 256k on Qwen3.8: hybrid architecture, only 16 of 64 layers carry full attention, so the KV cache is 8.6 GB at FP8.
37
👋 Bonjour Twitter ! Découvrez nos batteries portables et câbles multifonctions, parfaits pour les pros en déplacement et les amateurs de gadgets. 🔗 bim-computing.com #TechNews #Innovation #NomadeDigital #Gadgets #BIMComputing
151
🔌 Les utilisateurs adorent notre câble OmniLink ! ⭐ "Solide et de bonne qualité." - Client Amazon ⭐ "Idéal pour recharger mon iPhone en déplacement." - Client Amazon Disponible sur Amazon : amazon.fr/BIM-Computing-Adap… #Tech #OmniLink #BIMComputing #AvisClients #Amazon
26
🔋 Gardez vos appareils chargés en déplacement avec notre Powerbank 10 000mAh. Compacte, puissante et rapide. 🚀💼 Disponible sur notre site ! bim-computing.com/produit/po… #Tech #Gadget #Powerbank #BIMComputing
1
2
50
🔌 Notre câble OmniLink est désormais sur Amazon ! Plus besoin d'attendre, commandez le vôtre dès maintenant et profitez de la livraison rapide. 💫 amazon.fr/BIM-Computing-Adap… #Tech #Gadget #Amazon #OmniLink
18
🔌 OmniLink de @BIM_Computing : le câble 4-en-1 pour tous vos appareils ! Charge rapide, transfert de données efficace, et compatibilité universelle. 🌟 #Tech #Innovation #StayConnected bim-computing.com/produit/om…
1
3
35
🚀 Nouvelle Powerbank 10 000mAh de @BIM_Computing ! Ultra-légère, rapide et compatible avec tous vos appareils. Restez connecté en toutes circonstances. 🔋⚡📱 #Tech #Innovation #Gadget #StayConnected bim-computing.com/produit/po…
1
18