Conversation models that run on 256 GB

628 models with published weights that fit in 256 GB — a Mac Studio with 256. Counted at four-bit quantisation, weights only, leaving the machine about a third of its memory and a gigabyte for context: room for roughly 274 billion parameters. A mixture of experts is counted in full, because it is held in full even though only a few experts compute. 628 in all, a hundred to a page; this is page 4 of 7.

A Mac Studio at the top of its configuration, or a serious machine. Almost everything openly published fits, including the large mixtures of experts. At this point the constraint is no longer whether the model loads but whether it generates fast enough to be worth waiting for.

Models that answer in prose across whatever subject you put to them. This is the largest category and the least differentiated: most of them are competent at most things, so the choice usually comes down to price, context length and how the maker behaves about availability. Look at what a model costs per million tokens in and out — the output rate is often four or five times the input rate, and it is the one that decides your bill.

• dolphin-2.9-llama3-8bDolphin8.0B≈5.2 GB at 4-bit
• UserLM-8bMicrosoft8.0B≈5.2 GB at 4-bit
• Llama3-ChatQA-1.5-8BNVIDIA8.0B≈5.2 GB at 4-bit
• DarkIdol-Llama-3.1-8B-Instruct-1.2-Uncensoredaifeifei7988.0B≈5.2 GB at 4-bit
• KernelLLMfacebook8.0B≈5.2 GB at 4-bit
• Llama-3-8B-Instruct-Gradient-1048kDeepSky8.0B≈5.2 GB at 4-bit
• Llama-3-8B-WebMcGill NLP Group8.0B≈5.2 GB at 4-bit
• Llama-3.1-Nemotron-Nano-8B-v1NVIDIA8.0B≈5.2 GB at 4-bit
• Llama-3.1-SuperNova-LiteArcee AI8.0B≈5.2 GB at 4-bit
• Llama3-8B-Chinese-Chatshenzhi-wang8.0B≈5.2 GB at 4-bit
• Meta-Llama-3.1-8B-Instruct-abliteratedmlabonne8.0B≈5.2 GB at 4-bit
• llama-3-Korean-Bllossom-8BMLP-LAB8.0B≈5.2 GB at 4-bit
• aya-23-8BCohere Labs8.0B≈5.2 GB at 4-bit
• c4ai-command-r7b-12-2024Cohere Labs8.0B≈5.2 GB at 4-bit
• Molmo-7B-D-0924Ai28.0B≈5.2 GB at 4-bit
• LLaDA-8B-InstructGSAI-ML8.0B≈5.2 GB at 4-bit
• Apertus-8B-2509swiss-ai8.0B≈5.2 GB at 4-bit
• Apertus-8B-Instruct-2509swiss-ai8.0B≈5.2 GB at 4-bit1 also selling it hosted
• Apertus-v1.5-8Bswiss-ai8.0B≈5.2 GB at 4-bit1 also selling it hosted
• Bio-Medical-MultiModal-Llama-3-8B-V1ContactDoctor8.0B≈5.2 GB at 4-bit
• Foundation-Sec-8Bfdtn-ai8.0B≈5.2 GB at 4-bit
• Hermes 2 Pro Llama 3 8BNous Research8.0B≈5.2 GB at 4-bit2 also selling it hosted
• Idefics3-8B-Llama3HuggingFaceM48.0B≈5.2 GB at 4-bit
• L3 8B Stheno V3.2Sao10K8.0B≈5.2 GB at 4-bit4 also selling it hosted
• L3-8B-Lunaris-v1Sao10K8.0B≈5.2 GB at 4-bit2 also selling it hosted
• Llama-3-8B-Lexi-UncensoredOrenguteng8.0B≈5.2 GB at 4-bit
• Llama-3-ELYZA-JP-8Belyza8.0B≈5.2 GB at 4-bit
• Llama-3-Groq-8B-Tool-UseGroq8.0B≈5.2 GB at 4-bit
• Llama-3-Open-Ko-8Bbeomi8.0B≈5.2 GB at 4-bit
• Llama-3.1-8B-Lexi-Uncensored-V2Orenguteng8.0B≈5.2 GB at 4-bit1 also selling it hosted
• Llama-3.1-Nemotron-Nano-VL-8B-V1NVIDIA8.0B≈5.2 GB at 4-bit
• Llama-3.1-Storm-8Bakjindal532448.0B≈5.2 GB at 4-bit
• Llama-3.1-Tulu-3-8BAllen Institute for AI (Ai2)8.0B≈5.2 GB at 4-bit
• Llama3-OpenBioLLM-8Baaditya8.0B≈5.2 GB at 4-bit
• Llama3-TAIDE-LX-8B-Chat-Alpha1taide8.0B≈5.2 GB at 4-bit
• Meta-Llama-Guard-2-8BMeta Llama8.0B≈5.2 GB at 4-bit1 also selling it hosted
• MiniCPM4-8BOpenBMB8.0B≈5.2 GB at 4-bit
• Mistral-NeMo-Minitron-8B-BaseNVIDIA8.0B≈5.2 GB at 4-bit
• NeuralDaredevil-8B-abliteratedmlabonne8.0B≈5.2 GB at 4-bit
• WeDLM-8B-InstructTencent Hunyuan8.0B≈5.2 GB at 4-bit
• aya-vision-8bCohere Labs8.0B≈5.2 GB at 4-bit
• deepthought-8b-llama-v0.01-alpharuliad8.0B≈5.2 GB at 4-bit
• granite-3.1-8b-instructIBM Granite8.0B≈5.2 GB at 4-bit
• granite-3.3-8b-instructIBM Granite8.0B≈5.2 GB at 4-bit1 also selling it hosted
• Ling-3.0-tinyInclusionAI7.9B≈5.1 GB at 4-bit
• EXAONE-3.0-7.8B-InstructLG AI Research7.8B≈5.1 GB at 4-bit
• phixtral-4x2_8mlabonne7.8B≈5.1 GB at 4-bit
• EXAONE-3.5-7.8B-InstructLGAI-EXAONE7.8B≈5.1 GB at 4-bit
• internlm2_5-7b-chatinternlm7.7B≈5.0 GB at 4-bit
• Qwen-7B-ChatAlibaba7.7B≈5.0 GB at 4-bit
• Qwen1.5-7B-ChatAlibaba7.7B≈5.0 GB at 4-bit
• DeepHat-V1-7BDeepHat7.6B≈5.0 GB at 4-bit
• Marco-o1ATH-MaaS7.6B≈5.0 GB at 4-bit
• Qwen2.5-7B-Instruct-1MAlibaba7.6B≈5.0 GB at 4-bit
• VulnLLM-R-7BVirtueAI7.6B≈5.0 GB at 4-bit
• falcon-H1R-7BTechnology Innovation Institute7.6B≈4.9 GB at 4-bit
• llava-v1.6-mistral-7bliuhaotian7.6B≈4.9 GB at 4-bit
• starvector-8b-im2svgstarvector7.5B≈4.9 GB at 4-bit
• falcon-mamba-7bTechnology Innovation Institute7.3B≈4.7 GB at 4-bit
• Starling-LM-7B-alphaBerkeley-Nest7.2B≈4.7 GB at 4-bit
• Starling-LM-7B-betaNexusflow7.2B≈4.7 GB at 4-bit
• dolphin-2.1-mistral-7bDolphin7.2B≈4.7 GB at 4-bit
• dolphin-2.2.1-mistral-7bDolphin7.2B≈4.7 GB at 4-bit
• dolphin-2.8-mistral-7b-v02Dolphin7.2B≈4.7 GB at 4-bit
• openchat-3.5-0106openchat7.2B≈4.7 GB at 4-bit
• Mistral-7B-v0.1Mistral AI7.2B≈4.7 GB at 4-bit
• neural-chat-7b-v3-1Intel7.2B≈4.7 GB at 4-bit
• zephyr-7b-alphaHugging Face H47.2B≈4.7 GB at 4-bit
• zephyr-7b-betaHugging Face H47.2B≈4.7 GB at 4-bit1 also selling it hosted
• falcon-7bTechnology Innovation Institute7.2B≈4.7 GB at 4-bit
• falcon-7b-instructTechnology Innovation Institute7.2B≈4.7 GB at 4-bit
• bloom-7b1BigScience Workshop7.1B≈4.6 GB at 4-bit
• llava-1.5-7b-hfLlava Hugging Face7.1B≈4.6 GB at 4-bit
• DeciLM-7BDeci AI7.0B≈4.6 GB at 4-bit
• chameleon-7bfacebook7.0B≈4.6 GB at 4-bit
• ALLaM-7B-Instruct-previewHUMAIN7.0B≈4.6 GB at 4-bit
• AlphaMonarch-7Bmlabonne7.0B≈4.5 GB at 4-bit
• BELLE-7B-2MBelleGroup7.0B≈4.5 GB at 4-bit
• BioMistral-7BBioMistral7.0B≈4.5 GB at 4-bit
• Chinese-Llama-2-7bLinkSoul7.0B≈4.5 GB at 4-bit
• DeepSeek-R1-Distill-QWEN-7BDeepSeek7.0B≈4.5 GB at 4-bit3 also selling it hosted
• Dream-v0-Instruct-7BDream-org7.0B≈4.5 GB at 4-bit
• FastVLM-7BApple7.0B≈4.5 GB at 4-bit
• K2-Horizon-7BInstitute of Foundation Models7.0B≈4.5 GB at 4-bit
• Lily-Cybersecurity-7B-v0.2segolilylabs7.0B≈4.5 GB at 4-bit
• Llama-2-7bMeta Llama7.0B≈4.5 GB at 4-bit
• Llama-2-7b-chatMeta Llama7.0B≈4.5 GB at 4-bit
• Llama2-Chinese-7b-ChatFlagAlpha7.0B≈4.5 GB at 4-bit
• MiMo-7B-RLXiaomiMiMo7.0B≈4.5 GB at 4-bit
• MiMo-VL-7B-RLXiaomiMiMo7.0B≈4.5 GB at 4-bit
• Mistral-7B-Instruct-v0.1Mistral AI7.0B≈4.5 GB at 4-bit3 also selling it hosted
• Mistral-7B-Instruct-v0.2Mistral AI7.0B≈4.5 GB at 4-bit2 also selling it hosted
• Mistral-7B-OpenOrcaOpenOrca7.0B≈4.5 GB at 4-bit
• Mistral-Trismegistus-7Bteknium7.0B≈4.5 GB at 4-bit
• Molmo-7B-O-0924Allen Institute for AI (Ai2)7.0B≈4.5 GB at 4-bit
• NeuralBeagle14-7Bmlabonne7.0B≈4.5 GB at 4-bit
• NeuralHermes-2.5-Mistral-7Bmlabonne7.0B≈4.5 GB at 4-bit
• OLMoE-1B-7B-0924Allen Institute for AI (Ai2)7.0B≈4.5 GB at 4-bit
• OpenHermes-2-Mistral-7Bteknium7.0B≈4.5 GB at 4-bit1 also selling it hosted
• Openhermes2.5 Mistral 7Bteknium7.0B≈4.5 GB at 4-bit2 also selling it hosted

Wider