Framework Desktop (AMD Ryzen AI Max+ PRO 495, 192GB)

The Framework Desktop with the AMD Ryzen AI Max+ PRO 495 and 192GB of unified LPDDR5X answers one question: how large a model can you run on a desk without a data center? Framework's answer is 192GB of soldered memory on a 256-bit bus delivering 273 GB/s of bandwidth, paired with 16 Zen 5 CPU cores, a 40-core Radeon 8065S GPU, and an NPU rated at up to 55 TOPS. The DIY Edition is $6,799, a pre-built with a 2TB NVMe drive and Fedora pre-installed is $7,449, and pre-orders opened on 2026-09-30 with shipments starting in November.

This is prosumer and small-team hardware, not data center gear. The entire system draws 120W sustained (140W boost) from a 400W FlexATX supply, weighs 3.1kg, measures 96.8 x 205.5 x 226.1mm, and runs Linux. What it buys you is capacity. An RTX 5090 has 32GB of VRAM and roughly 6.5x the bandwidth. This machine has 192GB of it. For dense models above 70B parameters at aggressive quantization, and for mixture-of-experts models that are cheap to compute but enormous to store, capacity is the binding constraint, not raw FLOPS.

Framework says the 192GB configuration runs 300B-class open-weight models at 4-bit quantization, and specifically cites DeepSeek-V4.1-Flash at Q2 and MiMo V2.6 Flash RL at MXFP4 as working examples. Treat those as vendor claims until independent benchmarks land, but the underlying math holds: 192GB holds weights that no single consumer GPU, and few dual-GPU workstations, can touch.

• Memory capacity: 192GB LPDDR5X-8533, soldered, 256-bit bus, shared between CPU, GPU, and NPU.
• Memory bandwidth: 273 GB/s. This number governs token generation for bandwidth-bound decode.
• NPU: up to 55 TOPS INT8. Useful for ONNX Runtime, Ryzen AI, and vision pipelines. Most LLM work lands on the iGPU via ROCm or Vulkan.
• GPU: Radeon 8065S, 40 RDNA 3.5 cores. Far weaker in raw FLOPS than a discrete card.
• CPU: 16 Zen 5 cores / 32 threads, 3.1GHz base, 5.2GHz boost, 64MB L3 cache.
• Power: 120W sustained, 140W boost, 400W FlexATX PSU (80 Plus Gold at 110V, Silver at 230V).
• Storage: two PCIe 4.0 x4 M.2 2280 sockets with heatspreaders, up to 8TB each.

Decode throughput scales roughly with bandwidth divided by bytes read per token. At 273 GB/s and a realistic 55 to 70 percent efficiency, a 70B model at 4-bit (roughly 40GB of weights) lands in the 4 to 7 tokens/second range. A 32B model at 4-bit reaches the low teens. Small 8B models are compute-limited rather than bandwidth-limited, so expect tens of tokens per second, capped by the Radeon 8065S rather than by memory.

Prefill is the opposite problem. Prompt processing is compute-bound, and 40 RDNA 3.5 cores will chew through long contexts noticeably slower than a discrete GPU. Long-context work is viable because you have memory for the KV cache, but time-to-first-token runs into seconds on large prompts.

For context: NVIDIA's DGX Spark matches the 273 GB/s but tops out at 128GB. Apple's M3 Ultra Mac Studio offers 819 GB/s and up to 512GB at a much higher price point, without x86 Linux. An RTX 5090 delivers roughly 1,792 GB/s but only 32GB.

• Llama 3.1 8B and Mistral 7B at Q8 or FP16, tens of tokens per second.
• Qwen 2.5 32B and Qwen 3 32B at Q4_K_M. This is the sweet spot for quality per token on this machine.
• Mixtral 8x7B at Q4, roughly 24GB of weights, fast because only two experts activate per token.
• Gemma 3 27B, Llama 3.2 Vision 11B, and other multimodal models, with room left over for image tokens.
• gpt-oss-120B style MoE models at MXFP4, which the 192GB pool handles with headroom for context.

• Llama 3.3 70B at Q4_K_M, around 40GB, single-digit tokens per second. Step up to Q5 or Q6 if you want the quality bump.
• DeepSeek-R1 (671B MoE, 37B active) at 1.58-bit or 2-bit dynamic quantization, the configuration that made 128GB machines famous. The extra 64GB here buys real context headroom instead of a squeezed KV cache.
• 200B-class dense models at 2 or 3-bit. Quality loss is real, but the model runs.
• Framework's cited examples: DeepSeek-V4.1-Flash at Q2 and MiMo V2.6 Flash RL at MXFP4.

• Llama 3.1 405B at 4-bit is roughly 203GB of weights before any KV cache, so it exceeds the pool. You would be forced to 2 or 3-bit, which most practitioners avoid for reasoning workloads.
• Anything large at full precision. FP16 70B is 140GB and leaves almost nothing for context.

Q4_K_M (or MXFP4 on models trained for it) is the right default. Quality degradation is small and weight size drops enough that context and KV cache still fit. Drop to Q3 or Q2 only when capacity forces it, and expect measurable loss on reasoning tasks. For 30B-class models, Q6_K is often worth the extra memory because there is room.

• Local LLM hobbyists who want to run 70B-to-120B models without a multi-GPU tower. This is close to the cheapest single-box path to that capability.
• Developers building agentic workflows who need a shared inference box for LangGraph, CrewAI, or custom tool-calling loops. Capacity matters more than raw speed when several model roles run concurrently.
• Small teams running internal inference on sensitive data. 273 GB/s is slow by GPU standards, but it is local, quiet, 120W, and far cheaper to operate than renting an H100 by the hour.
• Edge and field deployments where a 4.5L, 3.1kg box that runs off a standard outlet beats a rack.
• Fine-tuning: LoRA and QLoRA on models up to roughly 30B are workable. Full-parameter training is not. Buy this for inference.

Connectivity supports the agent use case: dual USB4 Type-C, dual USB-A 3.2 Gen 2, HDMI 2.1, two DisplayPort 2.1 (UHBR10), 5Gbit Ethernet, Wi-Fi 7, a 3.5mm combo jack, and two front Expansion Card slots.

vs NVIDIA DGX Spark (128GB): Same 273 GB/s bandwidth, same unified-memory philosophy, same target workload. The Spark ships with CUDA and the NVIDIA software stack, which is a real advantage if you are already in that ecosystem. The Framework gives you 64GB more memory, x86 Linux, swappable storage, and no runtime lock-in. If your models fit in 128GB, the Spark is the easier buy. If they do not, this is the only option near the price.

vs Mac Studio M3 Ultra: Apple's machine offers 819 GB/s and up to 512GB of unified memory, and it will generate tokens several times faster on models that fit. It is also more expensive once you configure large memory, runs macOS, and shuts you out of ROCm and the wider x86 Linux tooling. Choose the Mac for raw speed on Apple-supported runtimes. Choose the Framework for Linux, x86 compatibility, and a lower entry price at the 192GB tier.

vs dual RTX 5090: Two 5090s cost less than this system and deliver several times the compute and bandwidth, but only 64GB of combined VRAM. They will crush 30B models and choke on 120B MoE. Different tools for different constraints.

Pick the Framework Desktop when your bottleneck is how much model fits, not how many tokens per second it emits.