LFM2.5 2.6B on M5 Max · 64 GiB

Experiment Complete Published

Rapid-MLX 1.0.2 on M5 Max · 64 GiB, Metal

LFM2.5 2.6B as mlx-affine-8bit-g64 (8-bit), archive 256 input / 128 output: a 256-token prompt and 128 tokens out. Measured 12 Nov 2026, 22:28 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: litert-lm default, f16, model-dtype. The experiments here use different 8-bit schemes: litert-int8, Q8_0, mlx-affine-8bit-g64.

Experiment id s3-lfm2.5-2.6b-8-bit-rapid-mlx-256

Summary

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

MetricMedianRangeMean ± SDQuartiles
Decode speed tok/s, higher is better162.6 161.1–163.6 162.5 ± 0.8 162.4–162.9
Time to first token ms, lower is better61.2 58.9–61.9 60.7 ± 1.4 59.5–61.7
Peak memory GiB, lower is better3.28 3.23–3.41 3.29 ± 0.08 3.23–3.33
Prefill speed tok/s, higher is better4,220 4,175–4,389 4,261 ± 97.9 4,187–4,344

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
161.54,204256 tok in 60.9 ms, engine161.1127 tok in 788 ms, engine3.23nominal → nominal
261.84,182256 tok in 61.2 ms, engine162.9127 tok in 780 ms, engine3.33nominal → nominal
361.94,175256 tok in 61.3 ms, engine163.5127 tok in 777 ms, engine3.41nominal → nominal
458.94,389256 tok in 58.3 ms, engine162.8127 tok in 780 ms, engine3.23nominal → nominal
561.04,236256 tok in 60.4 ms, engine162.4127 tok in 782 ms, engine3.33nominal → nominal
659.04,380256 tok in 58.4 ms, engine162.4127 tok in 782 ms, engine3.23nominal → 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
Rapid-MLX
Maintainer
Rapid-MLX contributors
License
Apache-2.0, open source
Version
1.0.2
Commit
f1aa011aa91bb38140ed3fb68f6267b77447ee0f
Binary SHA-256
d5c9b638732cda0b2e2a816aa85cc1780e0a3c1b690bb23c8d9b4d3ac81a527c
Built from
package
Build flags
None
Backend
Metal
Threads
Not recorded
Interface
python-worker
Settings in effect
backend=Device(gpu, 0) model_source=artifact folder context_length=4096 batch_size=1 max_num_seqs=1 max_concurrent_requests=1 prefill_batch_size=1 completion_batch_size=1 prefill_step_size=2048 prefix_cache=false disk_caches=false speculative=false spec_decode=none suffix_decoding=false response_cache_entries=0 pflash=off kv_cache_dtype=bf16 kv_quantized=false admission_window_ms=0 stream_interval=1 gpu_memory_utilization=0.9 greedy=true temperature=0 sampler=argmax trust_remote_code=false telemetry=false weight_bits=8 weight_group_size=64

Model and weights

Base model
LFM2.5 2.6B (LFM2.5, 2.6B)
Upstream
LiquidAI/LFM2.5-2.6B@7141b56cd4650af95c7c7b86983443aed050c96a
Weights
LiquidAI/LFM2.5-2.6B-MLX-8bit@b372ebbb518c0e81617e25d8824427dd9ee1f08c
Format
safetensors
File 1
README.md · 2,648 bytes · sha256 0e60b04b3ef5ab4e03e971537ec84cd45256c82f00efce24095ba73bf3f467f0
File 2
chat_template.jinja · 5,496 bytes · sha256 aadd425edb1d42fe9288e5de3624c21c1c03099129b9cfea5978f1bc67222c0a
File 3
config.json · 1,952 bytes · sha256 6f0d4974b6a17ff0d1140196b89d56107fdb0b88b0a548ce3035c9d99f139c52
File 4
generation_config.json · 306 bytes · sha256 e08c9a83f4bfe64468c9671a2b69b60549728e2a9481387f3555471f64e958bc
File 5
model.safetensors · 2.67 GiB · sha256 04f7c6997b1501ae2551109aee50d57419f651a80d1a68a58a2c4d53202ee1ca
File 6
model.safetensors.index.json · 41,983 bytes · sha256 e18c6098614fe191dd3991c4403ee60830179e3d61c518496549b3890db9140c
File 7
tokenizer.json · 17.1 MiB · sha256 695be7802a0e4b8a81048f0ff5ebb7fc811a0ba5a6be63dbb24deb5a81096f41
File 8
tokenizer_config.json · 363 bytes · sha256 11f1de897317b489dd09199284528382eeb17f566398b16e2039bb2266d26b09
Quantization
mlx-affine-8bit-g64 (8-bit)
Weight bits
8
Group size
64
KV cache
model-dtype
Conversion
Published upstream by LiquidAI
Conversion command
Not recorded

Workload

Workload
Archive 256 input / 128 output
Prompt
256 tokens, from prompts/source.txt
Prompt SHA-256
c10d1b7bef62d95322f8f8d229dc66118d9375107bdc49b06f76287d058b2f6c
Chat template
ea663864491de7ade391839479860ca95541f892f72665c73251fbd4643b1bef
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
12 Nov 2026, 22:28 UTC
Harness
xybrid-bench 0.4.0 (777cb3d9fecaa7cc1a30694d98492c33dbc6d693)
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
8
Check arithmetic
Passed · sha256 68ce25cb2c4b530f85434cd352428b022149c641163300e566f4f9ee9d789bcf
Check capital
Passed · sha256 b1098d691672b55e98767ac84c6d8375f7bcc0d78154c861ea70ac2ff66e0ba7
Check opposite
Passed · sha256 7c90b06eef25984adf77ec226311829bbc1065a033e9ad6b26d9bd0dfdc65f95
Check weekday
Passed · sha256 03db656ee680fb1053be9c98631722488bb10269075159dbae5b000f0359dae2
Check color
Failed · sha256 a4849893c59e4aa15ff645cadd4cce0fbd882097352a75d8e6f23619578a73c6
Check sort
Failed · sha256 6904619ddb316a20bc1220a9088da9a4181f1f16766b6fdd188defd9cddc4c75
Check animal
Passed · sha256 3e9114b6df29561433c18b3a6abce2c635fbd8e9ff25b5fcee93149bfea52859
Check word
Passed · sha256 d1bf31457072aaa4c116ea5819a644456232fdd87b15203d3eafdd5e541f4e61
Check copy
Passed · sha256 d8c711dfa5fef68d1d0f86320daeca065b91bc8d0dbcea210f8b9ef80eefc6c2
Check needle-start
Passed · sha256 56a69645d0ff053b3b7001c224f07a374bf963a22c9fa4706dda5e56e3616870
Check needle-end
Failed · sha256 1a8ad5c1a3264694da79d5312eeb8f20a35e3c5522e4cd33bbbe072f6e724877
Check continuation-128
Passed · sha256 35b435d7f1c75bb05c07d58cad000525a1d62d6dbe7fd00ae7416953dfa26286
Detail
9/9 of the checks the model passes at BF16; needs 8 and the mandatory ones (needle-start, continuation-128). Left out, failed by every BF16 configuration: color, sort, needle-end.

Reproduce it

Commands
pnpm bench run --plan plans/s3-lfm2.5-2.6b-8-bit.json

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

Raw data