Qwen3.8 27B on M5 Max · 64 GiB

Experiment Complete Published

MLX LM 0.32.2 on M5 Max · 64 GiB, Metal

Qwen3.8 27B as mlx-affine-4bit-g64 (4-bit), archive 2048 input / 128 output: a 2048-token prompt and 128 tokens out. Measured 10 Nov 2026, 00:07 UTC by Glenn Sonna, Xybrid.

Eligible for ranking when the objective's required metrics are measured and this configuration is selected.Approximate comparison: The experiments use different effective KV cache precision: f16, model-dtype. Its weights are 4-bit in another scheme: mlx-affine-4bit-g64, where most use Q4_K_M.

Experiment id s2-qwen3.8-27b-4-bit-mlx-2048

Summary

Over 6 valid trials of 6 attempted (6 planned), computed from the raw trials below.

MetricMedianRangeMean ± SDQuartiles
Decode speed tok/s, higher is better33.1 33.0–33.1 33.1 ± 0.0 33.0–33.1
Time to first token ms, lower is better2,145 2,052–2,288 2,165 ± 103 2,087–2,257
Peak memory GiB, lower is better17.87 17.86–17.87 17.87 ± 0.00 17.87–17.87
Prefill speed tok/s, higher is better956.0 895.4–998.6 948.1 ± 45.0 908.2–981.8

Trials, metric by metric

The line runs from the lowest to the highest valid trial, the dot is the median and each tick is one trial.

Every trial

Raw token counts and durations, with the speeds computed from them. Invalid trials stay in the record with their reason; they are left out of the summary, never silently dropped.

Every attempted trial
TrialTime to first token msPrefill tok/sDecode tok/sPeak memory GiBThermalNote
12,087981.92,048 tok in 2,086 ms, engine33.1127 tok in 3,842 ms, engine17.87nominal → nominal
22,275900.72,048 tok in 2,274 ms, engine33.0127 tok in 3,843 ms, engine17.87nominal → nominal
32,088981.22,048 tok in 2,087 ms, engine33.0127 tok in 3,843 ms, engine17.86nominal → nominal
42,202930.82,048 tok in 2,200 ms, engine33.1127 tok in 3,842 ms, engine17.87nominal → nominal
52,052998.62,048 tok in 2,051 ms, engine33.1127 tok in 3,838 ms, engine17.87nominal → nominal
62,288895.42,048 tok in 2,287 ms, engine33.1127 tok in 3,843 ms, engine17.87nominal → nominal

Caveats

  • Publisher cards identify the base model but do not always pin the exact export source checkpoint revision.
  • Engines may differ in quantization scheme, effective KV cache precision, kernels, and activations.

Configuration

Device

Device
Apple Mac17,6
Cohort
Apple silicon Macs (desktop)
SoC
Apple M5 Max
CPU
18-core Apple CPU
GPU
Not recorded
NPU
Not recorded
RAM
64.00 GiB
OS
macOS 27.2.0 (26B123)
Kernel
27.2.0
Power
mains
Low power mode
No
Performance mode
Not recorded

Engine

Engine
MLX LM
Maintainer
Apple (ml-explore)
License
MIT, open source
Version
0.32.2
Commit
Not recorded
Binary SHA-256
9d46803799f99d6fc7c35fff2d4357c79b1a25ae98cc63efeaaebeb50a7b5f33
Built from
package
Build flags
None
Backend
Metal
Threads
Not recorded
Interface
python-worker
Settings in effect
backend=Device(gpu, 0) prompt_cache=false speculative=false greedy=true temperature=0 sampler=argmax prefill_step_size=2048 context_length=4096 kv_quantized=false batch_size=1 trust_remote_code=false weight_bits=4 weight_group_size=64

Model and weights

Base model
Qwen3.8 27B (Qwen3.8, 27B)
Upstream
Qwen/Qwen3.8-27B@37561816e66b5b9153b1b64c10968c18c1e6f4a0
Weights
mlx-community/Qwen3.8-27B-4bit@10c35caafbb80f7dc6a7a432cdd11af10a6d4818
Format
safetensors
File 1
README.md · 632 bytes · sha256 860ff845b4746d0b3e1d1afa4f13aac691ef49c7165054cedfb199371568c4c2
File 2
chat_template.jinja · 8,952 bytes · sha256 c3cf9e34abf4f9e36c2d72165aa9c132d3e2a725b6c2586aaa3a8af9d7a81041
File 3
config.json · 4,932 bytes · sha256 14b65a0ee06517060a6bbd979bb1a8ff54e7b304b1a1f01d54344b88b8285e85
File 4
generation_config.json · 202 bytes · sha256 e70c136c1b78ddc1fb0905bac8e733a4dc448d4f852a5dd75143fffc70be550e
File 5
model-00001-of-00003.safetensors · 4.98 GiB · sha256 6cc1508e96fb5d0865dfd5753a79f4ec60651bf3e2a82844a7e8ae9c60528c0d
File 6
model-00002-of-00003.safetensors · 4.99 GiB · sha256 83f2a20ca8058f486a3634a27faf99587f4cd3c156a83dee34fb99e6ac178670
File 7
model-00003-of-00003.safetensors · 4.99 GiB · sha256 31b8c91ef899f79efaaa69e3d2c096f6e2ebeb2ff20e29222abbd9ebc79e560a
File 8
model.safetensors.index.json · 218,281 bytes · sha256 13b840162b4cb35c66fef7df072f7dbb4717908204364f5e5d9f9655a2758fa8
File 9
preprocessor_config.json · 390 bytes · sha256 27225450ac9c6529872ee1924fcb0962ff5634834f817040f444118116f4e516
File 10
processor_config.json · 991 bytes · sha256 222e060084f093738ca7af4c89ef09d844e415876021325fdf7d5c6821b119d4
File 11
tokenizer.json · 19.1 MiB · sha256 06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523
File 12
tokenizer_config.json · 1,165 bytes · sha256 792fa3f0cb88b111e54ef3134c873531008c4df471d108da17903426e308aa7b
File 13
video_preprocessor_config.json · 385 bytes · sha256 7768af27c1fafa9cc9011c1dc20067e03f8915e03b63504550e11d5066986d13
File 14
vocab.json · 6.4 MiB · sha256 ce99b4cb2983d118806ce0a8b777a35b093e2000a503ebde25853284c9dfa003
Quantization
mlx-affine-4bit-g64 (4-bit)
Weight bits
4
Group size
64
KV cache
model-dtype
Conversion
Published upstream by mlx-community
Conversion command
Not recorded

Workload

Workload
Archive 2048 input / 128 output
Prompt
2,048 tokens, from prompts/source.txt
Prompt SHA-256
c10d1b7bef62d95322f8f8d229dc66118d9375107bdc49b06f76287d058b2f6c
Chat template
af16f654968d47d09a6c604e00370b12dbc21f76456f59c9e98b82e84d9f6375
Output
128 tokens, ignoring end of sequence
Context length
4,096
Batch size
1
Sampling
Greedy
Timing
One streamed request per trial
Warmup
2 runs, not counted
Trials planned
6
Cooldown
15 s between trials

Conditions

Ambient
Not recorded
Thermal at start
nominal
Thermal at end
nominal
Thermal gate
macOS thermal state nominal
Throttling detected
No

Provenance

Measured by
Glenn Sonna, Xybrid
Operator affiliation
Xybrid
Run at
10 Nov 2026, 00:07 UTC
Harness
xybrid-bench 0.4.0 (ccb0a8ac606e19454c16dd0b625e29509c726b72)
Methodology
v0.6
Publication
published
Published
16 Nov 2026, 10:23 UTC
Reviewed by
Glenn Sonna

Output check

Result
Passed
Method
sanity-v1-bf16-reference
Reference
checks/sanity-v1.json
Reference SHA-256
a702dee079525ae0ec90465b1c9e26052c99cb3e1c53e83dea05be4c194fc7ca
Required passing checks
11
Check arithmetic
Passed · sha256 6b51d431df5d7f141cbececcf79edf3dd861c3b4069f0b11661a3eefacbba918
Check capital
Passed · sha256 5dd272b4f316b776a7b8e3d0894b37e1e42be3d5d3b204b8a5836cc50597a6b1
Check opposite
Passed · sha256 f151cd9e12c8994edfa316e4624f956bc68460454869177f287146f8cc956a12
Check weekday
Passed · sha256 7d8af1de1262f150187d5938bf649d6c35726970685f916d1ec6392a801b9762
Check color
Passed · sha256 d486dfbd5fb578340bccbdd0a662527eab38550648d5f44517e5ac71b8824703
Check sort
Passed · sha256 a386b3c9b4b4786df5bb6474bab0a62b8476e2f3f9c8a6433aca40152840f6b7
Check animal
Passed · sha256 4b227777d4dd1fc61c6f884f48641d02b4d121d3fd328cb08b5531fcacdabf8a
Check word
Passed · sha256 3b635d3b4437e023e72a9df26201b6ab51285b376446446ebb379121994550f6
Check copy
Passed · sha256 d61e6890d0e3571b0c7693612d41f8d059a1fcdef4ba959c0af4144541d41907
Check needle-start
Passed · sha256 8b9af2e5b8b4cfddb87d171997b95c471f8d68a68e30fb5177abeb3332372edd
Check needle-end
Passed · sha256 e6589c736af32f81bf239c5eda6de41959122ccc1d05463b75e0eb0c70fc831c
Check continuation-128
Passed · sha256 ed26731fe8d39dacac77703998aa4bb784317ae565cb21a58f83f75d7e458142
Detail
12/12; needs 11 and every mandatory check (needle-start, needle-end, continuation-128); the model has no BF16 reference.

Reproduce it

Commands
pnpm bench run --plan plans/s2-qwen3.8-27b-4-bit.json

Three palindromic blocks per quantization class and workload; load, two warmups, one trial, unload per turn.

Raw data