Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
cache_source: string
cached_tokens: int64
completion_tokens: int64
concurrency: int64
decode_tps_client: double
decode_tps_server: double
e2e_s: double
error: null
finish_reason: string
model_reported: string
peak_rss_bytes: int64
phase: string
prefill_ms_server: double
prefill_tps_server: double
prompt_length_requested: int64
prompt_tokens: int64
scenario: string
target: string
trial: int64
ttft_ms_client: double
ttft_ms_server: double
verdict: struct<status: string, winner: string, wins: struct<omlx: int64, splash: int64, ollama: int64>, reas (... 201 chars omitted)
child 0, status: string
child 1, winner: string
child 2, wins: struct<omlx: int64, splash: int64, ollama: int64>
child 0, omlx: int64
child 1, splash: int64
child 2, ollama: int64
child 3, reasons: list<item: null>
child 0, item: null
child 4, comparisons: list<item: struct<length: int64, scenario: string, metric: string, values: struct<omlx: double, spla (... 64 chars omitted)
child 0, item: struct<length: int64, scenario: string, metric: string, values: struct<omlx: double, splash: double, (... 52 chars omitted)
child 0, length: int64
child 1, scenario: string
child 2, metric: string
child 3, values: struct<omlx: double, splash: double, ollama: double>
child 0, omlx: double
child 1, splash: double
child 2, ollama: double
child 4, winner: string
child 5, advantage: double
run_id: string
backends: list<item: struct<name: string, target: struct<name: string, base_url: string, model: string, family (... 113 chars omitted)
child 0, item: struct<name: string, target: struct<name: string, base_url: string, model: string, family: string, q (... 101 chars omitted)
child 0, name: string
child 1, target: struct<name: string, base_url: string, model: string, family: string, quant_class: string, quantizat (... 58 chars omitted)
child 0, name: string
child 1, base_url: string
child 2, model: string
child 3, family: string
child 4, quant_class: string
child 5, quantization: string
child 6, runtime: string
child 7, protocol: string
child 8, pid: null
child 2, port: int64
to
{'run_id': Value('string'), 'verdict': {'status': Value('string'), 'winner': Value('string'), 'wins': {'omlx': Value('int64'), 'splash': Value('int64'), 'ollama': Value('int64')}, 'reasons': List(Value('null')), 'comparisons': List({'length': Value('int64'), 'scenario': Value('string'), 'metric': Value('string'), 'values': {'omlx': Value('float64'), 'splash': Value('float64'), 'ollama': Value('float64')}, 'winner': Value('string'), 'advantage': Value('float64')})}, 'backends': List({'name': Value('string'), 'target': {'name': Value('string'), 'base_url': Value('string'), 'model': Value('string'), 'family': Value('string'), 'quant_class': Value('string'), 'quantization': Value('string'), 'runtime': Value('string'), 'protocol': Value('string'), 'pid': Value('null')}, 'port': Value('int64')})}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
cache_source: string
cached_tokens: int64
completion_tokens: int64
concurrency: int64
decode_tps_client: double
decode_tps_server: double
e2e_s: double
error: null
finish_reason: string
model_reported: string
peak_rss_bytes: int64
phase: string
prefill_ms_server: double
prefill_tps_server: double
prompt_length_requested: int64
prompt_tokens: int64
scenario: string
target: string
trial: int64
ttft_ms_client: double
ttft_ms_server: double
verdict: struct<status: string, winner: string, wins: struct<omlx: int64, splash: int64, ollama: int64>, reas (... 201 chars omitted)
child 0, status: string
child 1, winner: string
child 2, wins: struct<omlx: int64, splash: int64, ollama: int64>
child 0, omlx: int64
child 1, splash: int64
child 2, ollama: int64
child 3, reasons: list<item: null>
child 0, item: null
child 4, comparisons: list<item: struct<length: int64, scenario: string, metric: string, values: struct<omlx: double, spla (... 64 chars omitted)
child 0, item: struct<length: int64, scenario: string, metric: string, values: struct<omlx: double, splash: double, (... 52 chars omitted)
child 0, length: int64
child 1, scenario: string
child 2, metric: string
child 3, values: struct<omlx: double, splash: double, ollama: double>
child 0, omlx: double
child 1, splash: double
child 2, ollama: double
child 4, winner: string
child 5, advantage: double
run_id: string
backends: list<item: struct<name: string, target: struct<name: string, base_url: string, model: string, family (... 113 chars omitted)
child 0, item: struct<name: string, target: struct<name: string, base_url: string, model: string, family: string, q (... 101 chars omitted)
child 0, name: string
child 1, target: struct<name: string, base_url: string, model: string, family: string, quant_class: string, quantizat (... 58 chars omitted)
child 0, name: string
child 1, base_url: string
child 2, model: string
child 3, family: string
child 4, quant_class: string
child 5, quantization: string
child 6, runtime: string
child 7, protocol: string
child 8, pid: null
child 2, port: int64
to
{'run_id': Value('string'), 'verdict': {'status': Value('string'), 'winner': Value('string'), 'wins': {'omlx': Value('int64'), 'splash': Value('int64'), 'ollama': Value('int64')}, 'reasons': List(Value('null')), 'comparisons': List({'length': Value('int64'), 'scenario': Value('string'), 'metric': Value('string'), 'values': {'omlx': Value('float64'), 'splash': Value('float64'), 'ollama': Value('float64')}, 'winner': Value('string'), 'advantage': Value('float64')})}, 'backends': List({'name': Value('string'), 'target': {'name': Value('string'), 'base_url': Value('string'), 'model': Value('string'), 'family': Value('string'), 'quant_class': Value('string'), 'quantization': Value('string'), 'runtime': Value('string'), 'protocol': Value('string'), 'pid': Value('null')}, 'port': Value('int64')})}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Qwen3.8 27B: oMLX vs Splash vs Ollama
Reproducible local benchmark on a MacBook Pro M4 Max with 64 GB unified memory.
Result in one sentence
For this workload, Ollama is the best overall backend: it wins most TTFT/decode comparisons and the concurrency tests. Splash is interesting specifically for cold long-context prefill, where it is faster than oMLX and slightly faster than Ollama at 32K tokens.
Scope
- Prompt lengths: 1K, 4K, 8K, 16K, 32K tokens
- Generation cap: 128 tokens
- Repetitions: 5 per cell, plus 1 warmup
- Scenarios: cold, warm incremental, warm repeat
- Concurrency: 1, 2, and 4 where applicable
- Backends were started and stopped between runs, with isolation checks
Artifacts
report.md: human-readable medians and metric definitionssummary.json: aggregated medians and p05/p95 values per cellresults.jsonl: raw successful request recordsmanifest-public.json: public run metadata without local filesystem paths
Models
These are related Qwen3.8 27B 4-bit artifacts, but they are not bit-identical weights or quantization formats:
| Backend | Model | Quantization |
|---|---|---|
| oMLX | Qwen3.8-27B-oQ4e-mtp |
oQ4e |
| Splash | incoai/Qwen3.8-27B-Splash |
Splash-4bit |
| Ollama | qwen3.8:27b-mlx |
nvfp4 |
Therefore this is a backend-plus-artifact comparison, not a controlled comparison of one identical model file.
Important caveats
- Splash and oMLX did not expose server-side prefill time through the tested OpenAI-compatible API, so their prefill columns are unavailable. Their client TTFT and decode measurements remain usable.
- Peak RSS is the sampled process-tree resident memory, not total Apple unified-memory or Metal allocation. Splash's low RSS values should not be interpreted as total model memory usage.
- The Ollama logs emitted an
invalid option provided option=thinkwarning. The benchmark followed the existingbench_ollama.pyrequest shape, but a formal rerun should remove or relocate that option before publication as a definitive comparison. - Splash's unload endpoint returned 404, although its server process was stopped and the isolation checks passed.
Overall verdict
The automatic verdict is Ollama, with 13 primary wins; Splash has 2; oMLX has 0. This does not disprove every performance claim made for Splash. It narrows the supported claim on this machine and workload: Splash appears valuable for cold, long-context prefill, but not as the fastest general-purpose backend in this three-way run.
See report.md for the complete table and metric definitions.
- Downloads last month
- 19