How Much RAM Do You Need to Run the Best Local LLMs in July 2026?

Compared Qwen 3.6 OpenAI gpt-oss

Here is the July 2026 answer: 16GB is enough to run a genuinely useful current local model. 32GB of unified memory is the practical floor for the best 27B-to-35B models that still make sense on consumer hardware. On a PC, 24GB of dedicated VRAM cleanly covers 27B-to-31B models; 35B at about 4-bit is a better fit for 32GB VRAM. If local AI will be a daily workload, 64GB is the comfortable buy.

That does not mean 64GB runs every leading open-weight model. The newest 119B to 122B models need about 96GB for clean 4-bit deployment, and DeepSeek V4 Flash-class models live in 192GB-plus territory. Anyone claiming that a model's small number of active parameters makes its full weight file fit in a normal laptop is confusing compute with storage.

RAM Requirements by Model Class

These are practical single-user tiers, not bare model-download sizes. They reserve room for the operating system, runtime, KV cache, and a useful context window. We have not hands-on benchmarked every hardware/model combination; the sizing is based on official model cards, official weight files, and conservative runtime headroom.

RAM, VRAM, and Unified Memory Are Not Interchangeable

On an Apple Silicon Mac, unified memory is shared by the CPU and GPU. That makes a 32GB or 64GB Mac unusually flexible for local models, but macOS and your applications still take part of the same pool.

On a Windows or Linux PC, VRAM determines how much of the model runs at full GPU speed. System RAM can hold overflow layers, but CPU offload is much slower. A machine with 16GB VRAM and 32GB system RAM is not equivalent to a 48GB GPU. It has 48GB of storage capacity across two pools, not 48GB of equally fast model memory.

How the Memory Math Actually Works

A useful approximation for GGUF-style Q4_K_M files is about 0.6GB per billion parameters. Q8 is roughly 1GB per billion parameters, and BF16 is roughly 2GB per billion. Q4 is about half the size of Q8 and about 75% smaller than BF16 — not 75% smaller than Q8.

The weight file is only the start. You also need memory for the runtime, temporary buffers, the vision or audio components of multimodal models, and the KV cache that holds your conversation context. Advertised 128K or 256K context does not mean that context fits on consumer hardware.

8GB: It Runs, but Do Not Buy It for Local AI

Pick: Gemma 4 E2B at Q4

Eight gigabytes of shared system or unified memory is enough for compact edge models, basic summarization, document classification, and simple offline chat. Gemma 4 E2B is the safer fit. Google's official table estimates a 4.5GB Q4 load for E4B before software and KV cache, so E4B is better suited to an 8GB dedicated GPU backed by separate system RAM than to an 8GB unified-memory computer.

The weakness is not whether it launches. It is the capability ceiling and the lack of working headroom. If local AI is part of the reason you are buying a machine, skip 8GB.

16GB: The Minimum for a Genuinely Useful Local Model

Pick: Gemma 4 12B at Q4

Sixteen gigabytes is viable, not pointless. Google lists a 6.7GB Q4 load for Gemma 4 12B, including 20% loading overhead but excluding software and context. Google separately positions the model for laptops with 16GB of VRAM or unified memory. That leaves enough room for a sensible context window and a normal desktop, provided you are not also running a heavy creative or development stack.

Runner-up: gpt-oss-20b. OpenAI says its native low-bit release runs within 16GB of memory. Qwen 3.5 9B is another strong choice when you would rather keep more headroom for context and other applications.

The tradeoff is model class. Sixteen gigabytes does not cleanly unlock the latest 27B to 35B models. If you already own a 16GB machine, use it. If you are buying specifically for local AI, move up to 32GB.

24GB VRAM: Excellent, but Watch the Context Window

Pick: Gemma 4 31B at Q4

Google's official Q4 load estimate for Gemma 4 31B is 17.5GB before software and KV cache. That fits on a 24GB GPU and leaves enough space for a useful but not enormous context window. In practice, 8K to 16K context is the safer target when the whole model stays on the GPU.

Qwen 3.6 27B is another strong fit. Qwen 3.6 35B-A3B is tighter: Qwen's comparable official 3.5 35B-A3B GPTQ-Int4 checkpoint is roughly 24.5GB on disk before runtime or cache, so a 24GB card requires a smaller quantization or some CPU offload.

On a 24GB unified-memory Mac, remember that the operating system shares the pool. Gemma 4 26B-A4B is a less cramped choice than forcing a 31B or 35B model into the last available gigabyte.

32GB: The Minimum Worth Buying for Serious Local AI

Pick: Qwen 3.6 35B-A3B at about 4-bit

This is the first tier that cleanly unlocks the strongest current consumer-local class. Qwen 3.6 27B, Qwen 3.6 35B-A3B, and Gemma 4 31B all become practical without treating every open browser tab as a threat.

The important word is practical. Some of these models can technically load below 32GB, but the remaining memory determines whether you get a useful context window, whether the runtime stays stable, and whether you can keep your IDE or browser open.

For a new Mac, 32GB unified memory is the minimum serious local-AI configuration. For a PC, the closest equivalent is a 24GB GPU plus at least 32GB of system RAM. The PC will usually be faster when the whole model stays in VRAM; the Mac gives you a more flexible shared pool.

64GB: The Comfortable Consumer Sweet Spot

Pick: Qwen 3.6 35B-A3B or Gemma 4 31B with more precision and context

Sixty-four gigabytes is less about unlocking a totally different 2026 model class and more about removing compromises. You can run the best 27B to 35B consumer models at Q8 or another higher-quality quantization, give them much longer context, and keep a development environment, browser, and document stack open beside the model.

It is also a good home for older 70B-class models at Q4. Do not count gpt-oss-120b as a clean 64GB fit: its native files leave essentially no room for cache or runtime, and OpenAI targets a single 80GB GPU.

What 64GB does not do: it does not cleanly fit the newest 119B to 122B models at normal 4-bit precision. Their official low-bit checkpoints are already larger than 64GB before cache and runtime. Do not buy a 64GB machine assuming that every model labeled “10B active” or “6.5B active” will fit.

96GB: The First Clean 119B to 122B Tier

Pick: Qwen 3.5 122B-A10B at Int4

Qwen's official Int4 checkpoint is about 78.9GB before runtime and cache. Mistral Small 4 119B tells the same story: its official NVFP4 checkpoint is about 70.8GB. Both exceed a 64GB machine before a single prompt is loaded.

That makes 96GB the first clean tier for this class. It is enough for the low-bit weights, a useful context window, and basic runtime overhead. If you expect to run other memory-heavy tools at the same time, 128GB is the safer configuration.

128GB and 192GB+: Flagship Workstation Territory

128GB pick: Step 3.5 Flash

StepFun's official llama.cpp guidance is unusually specific: Step 3.5 Flash uses 111.5GB of Int4 weights plus roughly 7GB of runtime overhead, requires at least 120GB, and recommends 128GB of unified memory. This is a real new capability tier, not merely extra headroom.

192GB-plus example: DeepSeek V4 Flash. DeepSeek's official mixed FP4/FP8 checkpoint is roughly 160GB before runtime and context. A 192GB system is the practical starting point, and more memory is preferable. The full GLM-5 and DeepSeek V4 Pro class remains multi-GPU server territory even after quantization.

The MoE Trap: Active Parameters Do Not Set RAM Use

Mixture-of-experts models advertise two numbers. Qwen 3.5 122B-A10B has about 122B total parameters but activates about 10B for a token. The smaller active number helps inference compute and speed. It does not shrink the stored model to 10B.

Every expert weight still needs to live somewhere. The same rule applies to Qwen 3.6 35B-A3B, Mistral Small 4 119B-A6.5B, DeepSeek V4 Flash 284B-A13B, and GLM-5's roughly 744B/40B-active design. Size hardware from the total checkpoint, not the active parameter count.

Mac vs PC: What We Would Buy

Mac: 32GB unified memory is the minimum serious purchase. Choose 64GB if local AI will be used daily. Choose 96GB to enter the 119B to 122B model class, and 128GB if you specifically want Step 3.5 Flash or more working headroom.

PC: prioritize dedicated GPU memory. A 24GB GPU plus 32GB of system RAM is the value entry point for current 27B to 31B models. A 32GB GPU is a cleaner home for 35B-class models. For 70GB-plus checkpoints, plan multiple GPUs or workstation hardware rather than pretending slow CPU offload is the same experience.

CPU-only PC: capacity is not the same as speed. A 32GB or 64GB machine may load a quantized model, but ordinary desktop DDR memory will generate much more slowly than a modern GPU or high-bandwidth Apple unified memory.

What About the Absolute Best Open-Weight Models?

The absolute benchmark leader and the best model that is sensible on a consumer machine are different questions. GLM-5 has roughly 744B total parameters. DeepSeek V4 Pro is larger still. Quantization cannot turn those into 64GB laptop models.

If you need those exact checkpoints, use a managed API, rented multi-GPU server, or a serious inference workstation. Buying consumer hardware to avoid a modest API bill rarely makes financial sense at that scale.

Verified Sizing Sources

• Google's Gemma 4 memory table lists Q4 load estimates of 6.7GB for 12B, 14.4GB for 26B-A4B, and 17.5GB for 31B, excluding software and KV cache.
• OpenAI's gpt-oss announcement specifies 16GB for gpt-oss-20b and a single 80GB GPU for gpt-oss-120b.
• Qwen's official Qwen 3.6 27B card documents the current dense 27B model used for the 24GB-VRAM recommendation.
• Qwen's official Qwen 3.6 card documents the 35B-total/3B-active architecture. Qwen's comparable official Qwen 3.5 Int4 checkpoint is about 24.5GB before runtime and cache.
• StepFun's official Step 3.5 Flash card specifies 120GB minimum and 128GB unified memory recommended for llama.cpp.
• DeepSeek V4 Flash's model card documents 284B total parameters, 13B active parameters, and its large mixed-precision checkpoint.

The Verdict

If you already own 16GB, do not panic. Run Gemma 4 12B or another strong small model and see whether local AI earns a permanent place in your workflow.

If you are buying now, 32GB is the minimum worth buying for serious local AI in July 2026. It unlocks the strongest consumer-local 27B to 35B class. Choose 64GB if you want long context, higher precision, and room to work beside the model.

If “best” means the stronger 119B to 122B class, start at 96GB. Choose 128GB for Step 3.5 Flash. Even then, you still do not have enough memory for every frontier open-weight model.

FAQ

Is 16GB enough RAM for a local LLM in July 2026?

Yes. A 16GB machine can run a genuinely useful current model such as Gemma 4 12B at 4-bit quantization or gpt-oss-20b in its native low-bit format. It is not enough for the strongest 27B to 35B consumer-local models with comfortable context and operating-system headroom.

Is 32GB RAM enough for the best local LLMs?

Yes for the best models that are practical on consumer hardware. A 32GB unified-memory machine can run 4-bit versions of Qwen 3.6 27B, Qwen 3.6 35B-A3B, and Gemma 4 31B with useful context. On a PC, 24GB of dedicated VRAM cleanly covers the 27B to 31B class; a 35B model needs a smaller quantization, partial CPU offload, or a 32GB GPU.

Should I buy 32GB or 64GB for local AI?

Buy 32GB if cost matters and you are satisfied with 4-bit 27B to 35B models. Buy 64GB if this will be a daily local-AI workstation: it provides more context, higher-precision quantizations, room for coding tools and browsers, and fewer memory-pressure compromises.

How much memory do 119B to 122B local models need?

Current models such as Qwen 3.5 122B-A10B and Mistral Small 4 119B need roughly 70GB to 79GB for their official low-bit checkpoints before context, runtime, and operating-system overhead. Treat 96GB as the minimum clean tier and 128GB as the comfortable tier.

Do active parameters determine how much RAM an MoE model needs?

No. Active parameters mainly affect compute per token. The complete set of model weights still has to be stored. A 122B-A10B model activates about 10B parameters at a time but still needs memory for roughly 122B parameters.