Multimodal model by baidu · 423.53B parameters
Everything this chip runs, machine by machine: MacBook Pro 14" M5 Pro, MacBook Pro 16" M5 Pro, Mac mini M5 Pro
ERNIE 4.5 VL 424B A47B Base Paddle needs at least 145.9 GiB, more than the 20.6 GiB available for models, so it needs a Mac with more memory.
Specifications
Unvalidated forecasts for an 8K text context under a declared full-GPU scenario, at the lowest GPU bin sold with the chosen memory. Memory is reported as its components: exact artifact bytes and a KV allocation derived from a declared attention layout where the catalog carries them, a parameter proxy with no error bound where it does not, and the runtime's working memory, which a browser cannot measure. A fit is confirmed only when every component is known, so a model whose known memory fits reads "Fit unconfirmed" until ToolPiper measures it on your Mac, and one whose known memory alone exceeds the budget reads "Won't fit". Estimated speed ignores prompt length and prefill, and can overstate long-context performance. No measured speed error range is available. The recommendation score weighs declared preferences, not measured answer quality, and becomes a range where evidence is missing.
| Quantization | Memory | Speed | Verdict |
|---|---|---|---|
| Q8_0 | at least 414.2 GiB | ~0 tok/s · Estimated | Won't fit |
| Q6_K | at least 315.6 GiB | ~0 tok/s · Estimated | Won't fit |
| Q5_K_M | at least 268.2 GiB | ~1 tok/s · Estimated | Won't fit |
| Q4_K_M | at least 228.8 GiB | ~1 tok/s · Estimated | Won't fit |
| Q3_K_M | at least 189.3 GiB | ~1 tok/s · Estimated | Won't fit |
| Q2_K | at least 145.9 GiB | ~1 tok/s · Estimated | Won't fit |
Affiliate disclosure: the shop icons in this table are Amazon affiliate links. As an Amazon Associate, ModelPiper earns from qualifying purchases. Every fit rating, quantization and tok/s figure below is computed from the hardware specs and is not influenced by that.
Unvalidated forecasts for an 8K text context under a declared full-GPU scenario, at the lowest GPU bin sold with the chosen memory. Memory is reported as its components: exact artifact bytes and a KV allocation derived from a declared attention layout where the catalog carries them, a parameter proxy with no error bound where it does not, and the runtime's working memory, which a browser cannot measure. A fit is confirmed only when every component is known, so a model whose known memory fits reads "Fit unconfirmed" until ToolPiper measures it on your Mac, and one whose known memory alone exceeds the budget reads "Won't fit". Estimated speed ignores prompt length and prefill, and can overstate long-context performance. No measured speed error range is available. The recommendation score weighs declared preferences, not measured answer quality, and becomes a range where evidence is missing.
Closest matches by use case and parameter count.
ToolPiper downloads, manages, and runs models with one click. Apple Silicon optimized.
Get ToolPiper free