Dataset Viewer
Auto-converted to Parquet Duplicate
date
stringdate
2026-11-18 00:00:00
2026-11-18 00:00:00
pass
int64
1
3
model
stringclasses
1 value
backend
stringclasses
5 values
dtype
stringclasses
3 values
quant
float64
4
4
⌀
thinking
stringclasses
2 values
decoding
stringclasses
3 values
context
stringclasses
7 values
n_timed
float64
1
3
⌀
prompt_tok_s_median
float64
147
1.3k
⌀
gen_tok_s_median
float64
4.35
40.5
⌀
ttft_s_median
float64
0.4
111
⌀
spread_prompt_pct
float64
0
39.9
⌀
spread_gen_pct
float64
0
48.2
⌀
load_s
float64
0.97
7.03
mem_after_load_gb
float64
6.3
17.8
mem_driver_or_peak_gb
float64
6.3
18.5
rss_gb_after_load
float64
0.67
19.3
note
stringlengths
5
156
suite
stringclasses
1 value
env
stringclasses
3 values
2026-11-18
1
Cloudflare/clairic-flash
pytorch_mps
bfloat16
null
off (no thinking mode)
greedy
512
3
801.096971
12.900748
0.639124
34.172146
47.748497
6.282276
17.770693
18.531677
0.670456
pass1
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
1
Cloudflare/clairic-flash
pytorch_mps
bfloat16
null
off (no thinking mode)
greedy
4096
3
391.483356
7.911837
10.462769
1.852054
1.804564
6.282276
17.770693
18.531677
0.670456
pass1
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
1
Cloudflare/clairic-flash
pytorch_mps
bfloat16
null
off (no thinking mode)
greedy
16384
1
342.119965
6.470446
47.889634
0
0
6.282276
17.770693
18.531677
0.670456
pass1 deadline mid-length
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
1
Cloudflare/clairic-flash
pytorch_mps
bfloat16
null
off (no thinking mode)
greedy
32768
null
null
null
null
null
null
6.282276
17.770693
18.531677
0.670456
pass1 deadline
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
2
Cloudflare/clairic-flash
pytorch_mps
bfloat16
null
off (no thinking mode)
greedy
16384
3
636.866314
10.275276
25.725964
3.774753
0.367049
7.027654
17.770693
18.531677
0.682114
pass2_16k
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
2
Cloudflare/clairic-flash
pytorch_mps
bfloat16
null
off (no thinking mode)
greedy
512
3
796.080702
12.719919
0.643151
0.476989
0.803626
7.027654
17.770693
18.531677
0.682114
pass2_recheck
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
2
Cloudflare/clairic-flash
pytorch_mps
bfloat16
null
off (no thinking mode)
greedy
4096
3
791.307543
11.888517
5.176243
1.753835
1.887781
7.027654
17.770693
18.531677
0.682114
pass2_recheck
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
2
Cloudflare/clairic-flash
clairic_mlx
bfloat16
null
off (no thinking mode)
greedy
512
3
358.962185
14.165619
1.426334
0.503246
0.32421
0.972579
16.921317
16.921317
18.195419
mlx_bf16 metal active after load
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 134bf542, claimed commit 75c52b6
2026-11-18
2
Cloudflare/clairic-flash
clairic_mlx
bfloat16
null
off (no thinking mode)
greedy
4096
3
252.853729
8.969771
16.199089
39.889998
23.724441
0.972579
16.921317
16.921317
18.195419
mlx_bf16 metal active after load
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 134bf542, claimed commit 75c52b6
2026-11-18
2
Cloudflare/clairic-flash
clairic_mlx
bfloat16
null
off (no thinking mode)
greedy
16384
3
147.08227
4.34828
111.39344
8.76214
4.868412
0.972579
16.921317
16.921317
18.195419
mlx_bf16 metal active after load
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 134bf542, claimed commit 75c52b6
2026-11-18
2
Cloudflare/clairic-flash
clairic_mlx
mlx_4bit_backbone
4
off (no thinking mode)
greedy
512
3
357.827147
33.612689
1.430858
0.633134
1.761158
2.009965
15.559623
17.454154
19.252777
mlx_q4 metal active after load
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 134bf542, claimed commit 75c52b6
2026-11-18
2
Cloudflare/clairic-flash
clairic_mlx
mlx_4bit_backbone
4
off (no thinking mode)
greedy
4096
3
281.212745
18.461857
14.565485
11.013632
48.162042
2.009965
15.559623
17.454154
19.252777
mlx_q4 metal active after load
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 134bf542, claimed commit 75c52b6
2026-11-18
2
Cloudflare/clairic-flash
clairic_mlx
mlx_4bit_backbone
4
off (no thinking mode)
greedy
16384
3
200.925408
8.606177
81.542699
3.909375
7.977747
2.009965
15.559623
17.454154
19.252777
mlx_q4 metal active after load
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 134bf542, claimed commit 75c52b6
2026-11-18
2
Cloudflare/clairic-flash
pytorch_mps_systemone
bfloat16
null
off
official systemone (typed heads / xgrammar)
receipt_image
1
null
null
null
null
null
7.027654
17.770693
18.531677
0.682114
extract 9/9 in 45.36s, image/PDF primary, 0 misses
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
2
Cloudflare/clairic-flash
pytorch_mps_systemone
bfloat16
null
off
official systemone (typed heads / xgrammar)
invoice_pdf
1
null
null
null
null
null
7.027654
17.770693
18.531677
0.682114
extract 12/12 in 120.46s, image/PDF primary, 0 misses
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
2
Cloudflare/clairic-flash
pytorch_mps_systemone
bfloat16
null
off
official systemone (typed heads / xgrammar)
form_image
1
null
null
null
null
null
7.027654
17.770693
18.531677
0.682114
extract 13/13 in 43.97s, image/PDF primary, 0 misses
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx
bfloat16
null
off (no thinking mode)
greedy
512
3
1,295.443304
15.155407
0.395231
6.527019
3.310056
3.122865
17.771321
17.802571
18.028244
pass3 mlx_bf16 metal after load; fixed port 5e7db58
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx
bfloat16
null
off (no thinking mode)
greedy
4096
3
1,181.845599
15.301145
3.465766
1.034387
0.938102
3.122865
17.771321
17.802571
18.028244
pass3 mlx_bf16 metal after load; fixed port 5e7db58
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx
bfloat16
null
off (no thinking mode)
greedy
16384
3
996.093855
14.458953
16.448249
11.708711
16.365459
3.122865
17.771321
17.802571
18.028244
pass3 mlx_bf16 metal after load; fixed port 5e7db58
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx
mlx_4bit_backbone
4
off (no thinking mode)
greedy
512
3
1,232.183455
40.537378
0.415523
0.367569
0.127486
3.179521
7.147297
7.178551
3.555267
pass3 mlx_q4 metal after load; peak after extract 12.31GB, after 16K speed 29.02GB
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx
mlx_4bit_backbone
4
off (no thinking mode)
greedy
4096
3
1,089.906692
38.290529
3.75812
1.392214
0.799339
3.179521
7.147297
7.178551
3.555267
pass3 mlx_q4 metal after load; peak after extract 12.31GB, after 16K speed 29.02GB
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx
mlx_4bit_backbone
4
off (no thinking mode)
greedy
16384
3
1,130.515885
35.967093
14.492499
6.758959
3.170315
3.179521
7.147297
7.178551
3.555267
pass3 mlx_q4 metal after load; peak after extract 12.31GB, after 16K speed 29.02GB
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx_extract
bfloat16
null
off (no thinking mode)
clairic_mlx extract (vision + typed head)
receipt_image
1
null
null
null
null
null
3.122865
17.771321
17.802571
18.028244
pass3 mlx_bf16 image primary 0/9 in 21.21s, 0/9 vs PyTorch; text fallback 8/9 in 28.56s, 8/9 vs PyTorch; miss item_names size suffix
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx_extract
bfloat16
null
off (no thinking mode)
clairic_mlx extract (vision + typed head)
invoice_pdf
1
null
null
null
null
null
3.122865
17.771321
17.802571
18.028244
pass3 mlx_bf16 PDF primary 0/12 in 83.94s, 1/12 vs PyTorch; text fallback 12/12 in 54.11s, 12/12 vs PyTorch
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx_extract
bfloat16
null
off (no thinking mode)
clairic_mlx extract (vision + typed head)
form_image
1
null
null
null
null
null
3.122865
17.771321
17.802571
18.028244
pass3 mlx_bf16 image primary 1/13 in 37.91s (shirt_size only), 4/13 vs PyTorch; text fallback not run because primary scored above 0
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx_extract
mlx_4bit_backbone
4
off (no thinking mode)
clairic_mlx extract (vision + typed head)
receipt_image
1
null
null
null
null
null
3.179521
7.147297
12.306508
3.555267
pass3 mlx_q4 image primary 0/9 in 20.92s, 0/9 vs PyTorch; text fallback 8/9 in 26.21s, 8/9 vs PyTorch; miss item_names size suffix
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx_extract
mlx_4bit_backbone
4
off (no thinking mode)
clairic_mlx extract (vision + typed head)
invoice_pdf
1
null
null
null
null
null
3.179521
7.147297
12.306508
3.555267
pass3 mlx_q4 PDF primary 0/12 in 67.64s, 1/12 vs PyTorch; text fallback 12/12 in 50.67s, 12/12 vs PyTorch
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx_extract
mlx_4bit_backbone
4
off (no thinking mode)
clairic_mlx extract (vision + typed head)
form_image
1
null
null
null
null
null
3.179521
7.147297
12.306508
3.555267
pass3 mlx_q4 image primary 0/13 in 43.10s, 1/13 vs PyTorch; text fallback 13/13 in 41.93s, 13/13 vs PyTorch
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58
2026-11-18
3
Cloudflare/clairic-flash
clairic_mlx_convert
mlx_4bit_converted
4
off (no thinking mode)
null
null
null
null
null
null
null
null
1.110009
6.297297
6.297297
11.767441
pass3 convert() 3.53s wrote backbone.safetensors 6500.3MB; load converted dir 1.11s, metal peak 6.30GB, RSS 11.77GB; official example extract 7.36s no error
clairic-flash-2026-11-18
M5 Pro 64GB, macOS 27.0.1, mlx 0.32.4, clairic-mlx 0.1.0, tarball sha256 490f17dc, commit 5e7db58

Cloudflare Clairic-flash 9B on a 64GB M5 Pro Mac mini

Headline: I ran Cloudflare's Clairic-flash 9B open weights on my 64GB M5 Pro Mac mini. On the official PyTorch path it scored 34/34 fields on a test receipt, invoice PDF and form. My MLX port is faster, especially at 4-bit (40.5 tok/s at 512 context), and it scored 33/34 on text input. Image and PDF input on the MLX port is currently broken.

Status: MLX text input verified (33/34). MLX image and PDF input is currently broken, fix in progress. Use the official PyTorch path for images and PDFs (34/34 on this suite).

This is an independent local measurement from November 18, 2026. I did not upload any weights.

A background inference app was paused for these runs at the owner's request.

Clairic-flash speed on a 64GB M5 Pro Mac mini

Setup

  • Apple M5 Pro Mac mini, 64 GB unified memory, macOS 27.0.1, 20 GPU cores, AC power
  • PyTorch path: Python 3.12.15, torch 2.13.0, transformers 5.10.2, xgrammar 0.2.1, device=mps, dtype=bfloat16
  • MLX path (pass 3): mlx 0.32.4, Metal, official Clairic-flash weights loaded by the port
  • Speed settings: greedy, thinking off, 256 new tokens, 1 warmup + 3 timed runs, median reported

Speed

Values are median tok/s and time to first token. PyTorch headlines are from pass 2. MLX headlines are from pass 3, after the port fixes.

Backend Context Prompt tok/s Gen tok/s TTFT
MLX BF16 512 1295 15.2 0.40 s
MLX BF16 4K 1182 15.3 3.47 s
MLX BF16 16K 996 14.5 16.4 s
MLX 4-bit 512 1232 40.5 0.42 s
MLX 4-bit 4K 1090 38.3 3.76 s
MLX 4-bit 16K 1131 36.0 14.5 s
PyTorch BF16 (MPS) 512 796 12.7 0.64 s
PyTorch BF16 (MPS) 4K 791 11.9 5.18 s
PyTorch BF16 (MPS) 16K 637 10.3 25.7 s

Load time and memory

Backend Load Memory
MLX BF16 3.1 s 17.8 GB Metal peak
MLX 4-bit 3.2 s 7.18 GB Metal peak at load; 29.0 GB after the 16K run
MLX 4-bit convert output (6.5 GB) 1.1 s 6.3 GB Metal peak
PyTorch BF16 (MPS) 7.0 s 28.8 GB peak

convert --bits 4 wrote a 6.5 GB backbone plus the unchanged 261.5 MB head. That directory loaded in 1.1 s.

Extraction

Synthetic documents with known ground truth: a rendered receipt PNG, an invoice PDF and a registration form PNG. This is a 3-document field-level check. It is not Cloudflare's 576-field synthetic receipt eval, and it should not be presented as one.

Official PyTorch path (pass 2):

Task Input Time Score
Receipt image 45 s 9/9
Invoice PDF 121 s 12/12
Form image 44 s 13/13
Total 34/34

MLX port (pass 3):

  • Text input, 4-bit: 33/34 (receipt 8/9, a truncated item name; invoice 12/12; form 13/13). Matched PyTorch on 33/34 fields.
  • Image and PDF input: fails. BF16 scored 1/34 and 4-bit scored 0/34. Outputs were garbled (for example the receipt merchant came out as "bistrocof", and a subtotal of 16.00 as 160). The vision tower has not been numerically matched to the real Clairic vision weights.

pytest on Metal: 19/19 passed.

Lab claims

Cloudflare reports (self-reported, not reproduced here): fresh text fields 328/336; mixed-input 213/226; synthetic receipt fields 576/576. I did not run those evals.

Caveats

  • Three passes the same evening. Pass 1 was cut off after a 16K first run and skipped extraction. Pass 2 completed PyTorch extraction and a full 16K set. Pass 3 tested the fixed MLX port. 32K was never run.
  • Pass 1 4K PyTorch was about twice as slow as the pass 2 recheck, with tight internal spread. Treat pass 2 as the PyTorch headline.
  • Speed is backbone greedy generate, 256 new tokens, not the extraction decoder. Extraction has its own wall times above.
  • The official linear-attention fast path was not installed. Prefill and decode used the PyTorch implementation.
  • MLX BF16 16K decode spread was 16.4%. The median is 14.5 tok/s.
  • BF16 form text was not measured in pass 3 because the image primary scored 1/13 and the text fallback only ran on a 0 score or an error.
  • A background inference app was paused for each run at the owner's request, then restarted after each run.

Files

  • results.csv, REPORT.md, clairic-flash-banner.png
  • docs/: the synthetic receipt, invoice and form, plus ground truth
  • raw/: pass 1 PyTorch speed logs
  • raw/pass2/: pass 2 PyTorch extraction and speed, plus the older MLX port (extract crashed on bfloat16)
  • raw/pass3_compact.json: pass 3 MLX speed, memory and extraction scores after the port fixes. Detailed pass 3 logs stayed on the machine that ran the bench.

Wrapper scripts and logs that controlled the paused background app are not included.

Licenses and credit

Measurements and this card: CC BY 4.0, attribution Pratik Gandhi. The Clairic weights and official runtime are Apache-2.0. No model weights are redistributed.

Credit to Cloudflare and the Workers AI team.

Write-up on GitHub

How I ran this, what worked and what broke: Cloudflare Clairic-flash (MLX). It's part of Local AI field notes, where I log every model I run on my Mac mini.

The MLX port (code only) is at pratikgx/Clairic-Flash-MLX.

Write-up on GitHub

How I ran this, what broke, and the scripts: Cloudflare Clairic-flash (MLX). It's part of Local AI field notes, where I log every model I run on my Mac mini.

Downloads last month
15