amgadtewfik commited on
Commit
f34db9c
·
verified ·
1 Parent(s): 906f8bd

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +6 -5
README.md CHANGED
@@ -36,11 +36,12 @@ That's within reach of the M4's ~84 GB/s memory-bandwidth ceiling — these
36
  formats are now bandwidth-bound, not kernel-bound, on Apple hardware.
37
 
38
  ## Models
39
- | Model | File format | Prefill | Decode | Peak memory | Notes |
40
- |---|---|---:|---:|---:|---|
41
- | Qwen3.5-9B IQ4_NL | ATF v4-era | ~85 tok/s | 11.5 tok/s | 9.1 GB | Fused-GEMM prefill |
42
- | Qwen3.8-27B Q2_K_XL | ATF v4-era | not recorded | 5.75 tok/s | 9.8 GB | OOM guard used |
43
- | Qwen-AgentWorld-35B-A3B | v2 file, v5 reader | 0.1 tok/s | 4.67 tok/s | 11.57 GB | 256 experts, top-k 8 |
 
44
 
45
  *Benchmarks: Apple M4, 16 GB unified memory, greedy decode, 4k context, warm
46
  pass. All three run comfortably on a 16 GB Mac.*
 
36
  formats are now bandwidth-bound, not kernel-bound, on Apple hardware.
37
 
38
  ## Models
39
+ | Model | Size | Prefill | Decode |
40
+ |---|---|---|---|
41
+ | Qwen3.5-9B-IQ4_NL.atf | 5.52 GB | 65.00 tok/s (274 tok, 4.20s) | 14.04 tok/s (104 tok, 7.40s) |
42
+ | Qwen3.8-27B-UD-Q2_K_XL.atf | 9.52 GB | 18.90 tok/s (274 tok, 14.50s) | 4.97 tok/s (96 tok, 19.30s) |
43
+ | Qwen-AgentWorld-35B-A3B-UD-IQ2_M.atf | 11.64 GB | 24.30 tok/s (274 tok, 11.30s) | 6.93 tok/s (105 tok, 15.20s) |
44
+
45
 
46
  *Benchmarks: Apple M4, 16 GB unified memory, greedy decode, 4k context, warm
47
  pass. All three run comfortably on a 16 GB Mac.*