Best LLMs for 64 GB to 128 GB of Unified Memory

By LLM Hardware team · Methodology · Updated October 17, 2026

64 GB to 128 GB of memory is where 70B dense models and large mixture-of-experts models become possible on one desk. This tier is unified memory, not a discrete GPU: an Apple Mac Studio, an AMD Ryzen AI Max+ 395 (Strix Halo) mini PC, or an NVIDIA DGX Spark. By our VRAM formula (params x bytes x 1.2 + 2), a 70B at Q4 is 44 GB and at Q8 is 86 GB.

What fits at 64 GB, 96 GB and 128 GB

Leave room for the operating system: unified memory is shared with it, so a 64 GB machine does not give a model all 64 GB. AMD lists up to 128 GB of LPDDR5x-8000 on a 256-bit bus for the Ryzen AI Max+ 395, which works out to 256 GB/s. Qwen3-235B-A22B at Q4 needs 143 GB and only fits the 256 GB or 512 GB Mac Studio configurations.

Reflection Beam, published 16 October 2026, is 501.3B total and 23.4B active, Apache-2.0, text only, 1M context. Q4 is 302.8 GB, so 64 GB, 96 GB, and 128 GB all miss it. Unsloth’s UD-Q4_K_XL files summed to 306.5 GB on 17 October and miss a 256 GB machine. UD-IQ1_S summed to 103.1 GB: that file can sit on 128 GB and does not fit 96 GB. It is a 1-bit file, not Q4. ollama run beam is the cloud tag beam:cloud, not a local weight. Ollama listed nemotron-3.5-super:120b-a12b-q4_K_M at 90GB the same day, which is tighter than the 74 GB formula: 96 GB is the edge after macOS, and 128 GB fits that file. Q8 of Super was listed at 133GB and does not fit 128 GB. Size a different quant in the VRAM calculator. The 24 GB side of Bolt is on the Bolt hardware page, and the three machines are compared in DGX Spark vs Mac Studio vs Strix Halo.

Capacity versus bandwidth

Every one of these has more memory than any single consumer GPU, but less bandwidth than a 24 GB RTX 4090 (1,008 GB/s). A dense 70B reads all 44 GB of weights per token, so it is slow on a 256-273 GB/s box. MoE models with few active parameters, such as gpt-oss-120b (5.1B active), suit this tier best. The head-to-head is in DGX Spark vs Mac Studio vs Strix Halo.

When another option is better

• Models up to 32B: a 24 GB or 32 GB GPU is much faster for them. See best LLMs for 32 GB VRAM.
• One big MoE on a GPU you already own: llama.cpp can keep experts in system RAM. See the MoE offloading guide.
• Frontier-size models such as Qwen3-235B at Q4 (143 GB) and Muse Spark Open at Q4 (491.6 GB) need more than 128 GB. Unsloth’s 2-bit Muse Spark Open target is 256 GB. The 2-bit Nemotron 3.5 Bolt checkpoint is a 24 GB file, not a reason to buy this tier: see whether 24 GB actually runs Bolt.

Hardware for this tier

Check current price on Amazon. Mac Studio 64 GB and larger configurations are ordered from Apple; DGX Spark is listed on NVIDIA.

GMKtec EVO-X2 128 GB on Amazon

ASIN B0F53MLYQ6. Amazon Associates link. The live price is on that page, not locked in here.

Mac Studio M5 Max 36 GB on Amazon

ASIN B0HGKSQMX6. Amazon Associates link. The live price is on that page, not locked in here.

Sources

• Apple, Mac Studio technical specifications (checked 14 Oct 2026)
• AMD, Ryzen AI Max+ 395 specs (checked 14 Oct 2026)
• NVIDIA, DGX Spark specs (checked 14 Oct 2026)
• Llama 3.3 70B Instruct model card
• gpt-oss-120b model card (117B total, 5.1B active)
• Qwen3-235B-A22B model card (235B total, 22B active)
• NVIDIA, Nemotron 3.5 Super VL and Bolt (15 Oct 2026)
• Hugging Face, Muse Spark Open (816B total, about 42B active)
• LLMHardware methodology: the VRAM formula

A correction or a hardware question goes to the contact page.