Strix Halo vs a GPU for local AI
Strix Halo (Framework Desktop, Ryzen AI Max+ PRO 495) has 192 GB of unified memory at ~273 GB/s (computed from Framework's 256-bit LPDDR5X-8533 spec) vs 24 GB at 936 GB/s on an RTX 3090 (TechPowerUp). More fits, but each token is slower. (source)
| Spec | Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB | GeForce RTX 3090 |
|---|---|---|
| Memory | 192 GB LPDDR5X-8533 | 24 GB GDDR6X |
| Bandwidth | 273 GB/s | 936.2 GB/s |
| Low price | check price | low $1,889, avg $2,404 (used) source, 2026-10-15 |
| Qwen2.5-Coder-32B-Instruct estimate | ~10 tok/s | ~36 tok/s |
| $ per tok/s (Qwen2.5-Coder-32B-Instruct) | check price | $52 |
| Models that fit | 27 of 27 | 22 of 27 |
| Largest that fits | GLM-4.5-Air | Qwen2.5-Coder-32B-Instruct |
Last updated Oct 16, 2026. Speeds are bandwidth-based estimates, not benchmarks. Prices are dated snapshots and move daily; "check price" means we have no verified price.
What each one runs, model by model
| Model (Q4_K_M) | File | Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB estimate | GeForce RTX 3090 estimate |
|---|---|---|---|
| GLM-4.5-Air | 73.5 GB | ~26 tok/s | doesn't fit |
| gpt-oss-120b | 62.84 GB | ~76 tok/s | doesn't fit |
| Qwen2.5-72B-Instruct | 47.42 GB | ~4 tok/s | doesn't fit |
| Llama-3.3-70B-Instruct | 42.52 GB | ~5 tok/s | doesn't fit |
| DeepSeek-R1-Distill-Llama-70B | 42.52 GB | ~5 tok/s | doesn't fit |
| Qwen2.5-Coder-32B-Instruct | 19.85 GB | ~10 tok/s | ~36 tok/s |
| DeepSeek-R1-Distill-Qwen-32B | 19.85 GB | ~10 tok/s | ~36 tok/s |
| Qwen3-32B | 19.76 GB | ~10 tok/s | ~36 tok/s |
| Qwen3-30B-A3B | 18.63 GB | ~103 tok/s | ~353 tok/s |
| gemma-3-27b-it | 16.55 GB | ~13 tok/s | ~43 tok/s |
| Mistral-Small-3.2-24B-Instruct-2506 | 14.33 GB | ~14 tok/s | ~50 tok/s |
| Devstral-Small-2505 | 14.33 GB | ~14 tok/s | ~50 tok/s |
| gpt-oss-20b | 11.67 GB | ~103 tok/s | ~354 tok/s |
| Qwen3-14B | 9 GB | ~23 tok/s | ~79 tok/s |
| Qwen2.5-Coder-14B-Instruct | 8.99 GB | ~23 tok/s | ~79 tok/s |
| DeepSeek-R1-Distill-Qwen-14B | 8.99 GB | ~23 tok/s | ~79 tok/s |
| Mistral-Nemo-Instruct-2407 | 7.48 GB | ~28 tok/s | ~95 tok/s |
| gemma-3-12b-it | 7.3 GB | ~28 tok/s | ~97 tok/s |
| NVIDIA-Nemotron-Nano-9B-v2 | 6.53 GB | ~32 tok/s | ~109 tok/s |
| Qwen3-8B | 5.03 GB | ~41 tok/s | ~141 tok/s |
| Llama-3.1-8B-Instruct | 4.92 GB | ~42 tok/s | ~145 tok/s |
| Qwen2.5-Coder-7B-Instruct | 4.68 GB | ~44 tok/s | ~152 tok/s |
| Mistral-7B-Instruct-v0.3 | 4.37 GB | ~47 tok/s | ~163 tok/s |
| Qwen3-4B | 2.5 GB | ~83 tok/s | ~285 tok/s |
| gemma-3-4b-it | 2.49 GB | ~83 tok/s | ~286 tok/s |
| Llama-3.2-3B-Instruct | 2.02 GB | ~103 tok/s | ~352 tok/s |
| Qwen3-0.6B | 0.48 GB | ~432 tok/s | ~1482 tok/s |
File sizes are Q4_K_M GGUFs from Hugging Face (linked on each model page), accessed 2026-10-15. A model fits when its file plus 1.5 GB of headroom is at most the card's memory.
What the extra memory buys: 5 models fit the 192 GB Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB but not the 24 GB GeForce RTX 3090: GLM-4.5-Air (73.5 GB), gpt-oss-120b (62.84 GB), Qwen2.5-72B-Instruct (47.42 GB), Llama-3.3-70B-Instruct (42.52 GB), DeepSeek-R1-Distill-Llama-70B (42.52 GB).
Speed calculator estimate
Estimate = 0.76 × bandwidth ÷ GB read per token (Q4 file size, scaled by active/total parameters for MoE). Fit = file size + 1.5 GB headroom ≤ total VRAM. Multi-GPU: VRAM adds, speed uses one card's bandwidth (layer split). Short context only. How this works
Questions people ask
Which is faster on Qwen2.5-Coder-32B-Instruct?
On Qwen2.5-Coder-32B-Instruct (19.85 GB at Q4_K_M), the largest model that fits all of them: Framework Desktop, AMD Ryzen AI Max+ PRO 495 about 10 tok/s, GeForce RTX 3090 about 36 tok/s. Speed tracks memory bandwidth: Framework Desktop, AMD Ryzen AI Max+ PRO 495 273 GB/s, GeForce RTX 3090 936.2 GB/s.
Which is cheaper per token per second?
Lowest tracked price divided by the estimated speed on Qwen2.5-Coder-32B-Instruct: Framework Desktop, AMD Ryzen AI Max+ PRO 495 no verified price, GeForce RTX 3090 $52 per tok/s. Those prices are Framework Desktop, AMD Ryzen AI Max+ PRO 495 no verified price, GeForce RTX 3090 $1,889 (used), dated snapshots that move daily.
What fits on each of them?
Framework Desktop, AMD Ryzen AI Max+ PRO 495 (192 GB): 27 of our 27 sized models, largest GLM-4.5-Air at 73.5 GB; GeForce RTX 3090 (24 GB): 22 of our 27 sized models, largest Qwen2.5-Coder-32B-Instruct at 19.85 GB. Unified memory on the Apple and Framework systems is shared with the OS, so treat those fit results as optimistic.
Is the extra memory worth it?
5 models in our data fit the 192 GB Framework Desktop, AMD Ryzen AI Max+ PRO 495 but not the 24 GB GeForce RTX 3090: GLM-4.5-Air (73.5 GB), gpt-oss-120b (62.84 GB) and Qwen2.5-72B-Instruct (47.42 GB), and 2 more. Where both fit, on Qwen2.5-Coder-32B-Instruct, the estimates are Framework Desktop, AMD Ryzen AI Max+ PRO 495 about 10 tok/s, GeForce RTX 3090 about 36 tok/s.
Related
- What the Framework Desktop, AMD Ryzen AI Max+ PRO 495 (Strix Halo), 192 GB can run
- What the NVIDIA GeForce RTX 3090 can run
- What GPU you need for GLM-4.5-Air
- What GPU you need for Qwen2.5-Coder-32B-Instruct
- What GPU you need for Qwen3-32B
- Can the GeForce RTX 3090 run Qwen2.5-Coder-32B-Instruct?
- Can the GeForce RTX 3090 run Qwen3-32B?
- Build: 24 GB: one used RTX 3090 — $2,933.45, 24 GB
- Build: 48 GB: dual used RTX 3090 — $5,443.94, 48 GB
- Guide: RTX 3090 vs RTX 5090 for local LLMs
- Guide: Mac vs PC for local LLMs
- Guide: Can an RTX 5060 Ti 16 GB run Qwen3 Coder 30B?