ggml_metal_device_init: GPU name: MTL0 (Apple M5 Max)
ggml_metal_device_init: GPU family: MTLGPUFamilyApple10 (1010)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4 (5002)
ggml_metal_device_init: simdgroup reduction = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory = true
ggml_metal_device_init: has bfloat = true
ggml_metal_device_init: has tensor = true
ggml_metal_device_init: use residency sets = true
ggml_metal_device_init: use shared buffers = true
ggml_metal_device_init: recommendedMaxWorkingSetSize = 30150.67 MB
| model | size | params | backend | threads | fa | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --: | --------------: | -------------------: |
| gemma4 26B.A4B Q4_0 | 13.26 GiB | 25.23 B | MTL,BLAS | 6 | 1 | pp512 | 2999.32 ± 22.58 |
| gemma4 26B.A4B Q4_0 | 13.26 GiB | 25.23 B | MTL,BLAS | 6 | 1 | tg128 | 110.69 ± 0.24 |
build: 961e4b26a (9950)
ggml_metal_device_init: GPU name: MTL0 (Apple M5 Max)
ggml_metal_device_init: GPU family: MTLGPUFamilyApple10 (1010)
ggml_metal_device_init: GPU family: MTLGPUFamilyCommon3 (3003)
ggml_metal_device_init: GPU family: MTLGPUFamilyMetal4 (5002)
ggml_metal_device_init: simdgroup reduction = true
ggml_metal_device_init: simdgroup matrix mul. = true
ggml_metal_device_init: has unified memory = true
ggml_metal_device_init: has bfloat = true
ggml_metal_device_init: has tensor = true
ggml_metal_device_init: use residency sets = true
ggml_metal_device_init: use shared buffers = true
ggml_metal_device_init: recommendedMaxWorkingSetSize = 30150.67 MB
| model | size | params | backend | threads | test | t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | --------------: | -------------------: |
| gpt-oss 20B Q8_0 | 12.28 GiB | 20.91 B | MTL,BLAS | 6 | pp512 | 2867.67 ± 34.75 |
| gpt-oss 20B Q8_0 | 12.28 GiB | 20.91 B | MTL,BLAS | 6 | tg128 | 105.45 ± 0.11 |
build: 961e4b26a (9950)
curl -s localhost:11434/api/chat -d '{"model":"qwen3.5:9b","stream":false,"messages":[{"role":"system","content":"Datum: 2026-10-14 10:00. Ti si asistent."},{"role":"user","content":"Zdravo"}]}' | grep -o '"prompt_eval_cached_count":[0-9]*'
prompt_eval_cached_count padne na nulu. Stavite datum na kraj system prompta i ponovite: ostaje skoro ceo. Ovo sam uhvatio praveći alat za posmatranje lokalnih modela, LLMxRay https://github.com/LogneBudo/llmxray,npx llmxray, 

| Preis | 2.938,51 EUR | |
| AmazonGlobal-Versand | 85,07 EUR | |
| Geschätzte Importkosten | 641,01 EUR | |
| Summe | 3.664,58 EUR |
| Price | EUR 2,938.51 | |
| AmazonGlobal Shipping | EUR 57.20 | |
| Estimated Import Charges | EUR 0.00 | |
| Total | EUR 2,995.71 |
brzina mi nije bitna samo da može da učita Qwen 72B i da mu ostane lufta, malo mi fali da "povučem obarač"...na ebayu su 4000e minimum, nije mi jasno ovo "no import charges" kako oni mogu da odluče da li će kod nas naplatiti carinu ili ne ili oni refundiraju carinu ako se naplati
Follow along with the video below to see how to install our site as a web app on your home screen.
Napomena: this_feature_currently_requires_accessing_site_using_safari