Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed because of a cast error
Error code: DatasetGenerationCastError
Exception: DatasetGenerationCastError
Message: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 3 new columns ({'server', 'prefill_short_tps_median', 'prefill_long_tps'}) and 10 missing columns ({'engine', 'ttft_ms_median', 'app_version', 'app', 'condition', 'context', 'app_ram_mb_loaded', 'pre_load_vram_mib', 'cold_first_request_s', 'prefill_long_tps_median'}).
This happened while the csv dataset builder was generating data using
hf://datasets/iBlessi/local-llm-runtime-comparison-rtx-5080/llm-server-compare-2026-08.csv (at revision 884bc01db023c1c4e828c46263bd8f5cda71550c), ['hf://datasets/iBlessi/local-llm-runtime-comparison-rtx-5080@884bc01db023c1c4e828c46263bd8f5cda71550c/llm-server-compare-2026-08-26-lmstudio.csv', 'hf://datasets/iBlessi/local-llm-runtime-comparison-rtx-5080@884bc01db023c1c4e828c46263bd8f5cda71550c/llm-server-compare-2026-08.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1848, in _prepare_split_single
writer.write_table(table)
~~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 765, in write_table
self._write_table(pa_table, writer_batch_size=writer_batch_size)
~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 773, in _write_table
pa_table = table_cast(pa_table, self._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
model: string
server: string
decode_tps_median: double
decode_tps_wall_median: double
prefill_short_tps_median: double
prefill_long_tps: double
prefill_long_n_prompt: double
vram_total_mib_median: double
load_s: double
iterations: double
decode_tps_all: string
-- schema metadata --
pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 1710
to
{'model': Value('string'), 'app': Value('string'), 'app_version': Value('string'), 'engine': Value('string'), 'condition': Value('string'), 'context': Value('int64'), 'decode_tps_median': Value('float64'), 'decode_tps_wall_median': Value('float64'), 'prefill_long_tps_median': Value('float64'), 'prefill_long_n_prompt': Value('int64'), 'ttft_ms_median': Value('float64'), 'vram_total_mib_median': Value('int64'), 'pre_load_vram_mib': Value('int64'), 'load_s': Value('float64'), 'cold_first_request_s': Value('float64'), 'app_ram_mb_loaded': Value('float64'), 'iterations': Value('int64'), 'decode_tps_all': Value('string')}
because column names don't match
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1850, in _prepare_split_single
raise DatasetGenerationCastError.from_cast_error(
...<4 lines>...
)
datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
All the data files must have the same columns, but at some point there are 3 new columns ({'server', 'prefill_short_tps_median', 'prefill_long_tps'}) and 10 missing columns ({'engine', 'ttft_ms_median', 'app_version', 'app', 'condition', 'context', 'app_ram_mb_loaded', 'pre_load_vram_mib', 'cold_first_request_s', 'prefill_long_tps_median'}).
This happened while the csv dataset builder was generating data using
hf://datasets/iBlessi/local-llm-runtime-comparison-rtx-5080/llm-server-compare-2026-08.csv (at revision 884bc01db023c1c4e828c46263bd8f5cda71550c), ['hf://datasets/iBlessi/local-llm-runtime-comparison-rtx-5080@884bc01db023c1c4e828c46263bd8f5cda71550c/llm-server-compare-2026-08-26-lmstudio.csv', 'hf://datasets/iBlessi/local-llm-runtime-comparison-rtx-5080@884bc01db023c1c4e828c46263bd8f5cda71550c/llm-server-compare-2026-08.csv']
Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
model string | app string | app_version string | engine string | condition string | context int64 | decode_tps_median float64 | decode_tps_wall_median float64 | prefill_long_tps_median float64 | prefill_long_n_prompt int64 | ttft_ms_median float64 | vram_total_mib_median int64 | pre_load_vram_mib int64 | load_s float64 | cold_first_request_s float64 | app_ram_mb_loaded float64 | iterations int64 | decode_tps_all string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
gpt-oss-20b | lms | LM Studio 0.4.21 | llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1 | defaults | 8,192 | 186.26 | 184.34 | 5,241.3 | 474 | 78.9 | 12,310 | 1,119 | 8.43 | 0.36 | 12,917.9 | 5 | 183.42|186.26|186.84|186.98|185.4 |
gpt-oss-20b | lms | LM Studio 0.4.21 | llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1 | normalized | 4,096 | 242.83 | 240.13 | 9,617.7 | 474 | 38.6 | 12,533 | 1,101 | 8.43 | 0.27 | 13,176.6 | 5 | 242.83|242.82|243.14|241.74|244.35 |
gpt-oss:20b | ollama | ollama 0.32.15 | vendored llama.cpp 9d77fa172 | defaults | 4,096 | 213.54 | 212.03 | 9,239.4 | 475 | null | 13,570 | 1,053 | 18.56 | 23.67 | 2,138.8 | 5 | 213.22|213.64|213.48|213.54|213.76 |
gpt-oss:20b | ollama | ollama 0.32.15 | vendored llama.cpp 9d77fa172 | normalized | 4,096 | 213.65 | 211.8 | 9,187.1 | 474 | null | 13,570 | 1,053 | 8.05 | 8.36 | 2,152.1 | 5 | 213.62|213.65|213.68|213.25|213.73 |
llama-3.2-3b-q4km | lms | LM Studio 0.4.21 | llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1 | defaults | 8,192 | 314.86 | 308.68 | 20,341.8 | 440 | 9.4 | 4,260 | 1,122 | 2.75 | 0.07 | 3,471.9 | 5 | 273.28|312.54|316.75|315.22|314.86 |
llama-3.2-3b-q4km | lms | LM Studio 0.4.21 | llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1 | normalized | 4,096 | 316 | 309.91 | 20,336.7 | 442 | 9.2 | 3,808 | 1,103 | 2.76 | 0.07 | 3,469.6 | 5 | 310.55|316.0|315.86|316.23|316.6 |
llama3.2:3b | ollama | ollama 0.32.15 | vendored llama.cpp 9d77fa172 | defaults | 4,096 | 351.22 | 345.19 | 17,802.7 | 434 | null | 3,760 | 1,053 | 2.67 | 2.79 | 1,234 | 5 | 351.15|350.9|351.86|351.32|351.22 |
llama3.2:3b | ollama | ollama 0.32.15 | vendored llama.cpp 9d77fa172 | normalized | 4,096 | 350.52 | 338.11 | 17,862.3 | 432 | null | 3,760 | 1,057 | 2.27 | 2.38 | 1,239.6 | 5 | 349.35|350.38|350.52|350.63|350.96 |
qwen2.5-14b-q4km | lms | LM Studio 0.4.21 | llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1 | defaults | 8,192 | 98.15 | 97.31 | 5,195.8 | 454 | 24.6 | 11,194 | 1,103 | 7.38 | 0.15 | 10,406.8 | 5 | 97.36|98.14|98.17|98.15|98.17 |
qwen2.5-14b-q4km | lms | LM Studio 0.4.21 | llama.cpp-win-x86_64-nvidia-cuda12-avx2@2.29.1 | normalized | 4,096 | 97.63 | 96.62 | 5,566.5 | 453 | 24.7 | 10,405 | 1,106 | 7.68 | 0.16 | 10,411.1 | 5 | 96.89|97.75|97.46|97.79|97.63 |
qwen2.5:14b | ollama | ollama 0.32.15 | vendored llama.cpp 9d77fa172 | defaults | 4,096 | 102.25 | 101.06 | 4,983.8 | 453 | null | 10,356 | 1,053 | 5.51 | 5.67 | 1,604.1 | 5 | 102.13|102.34|102.25|102.3|102.25 |
qwen2.5:14b | ollama | ollama 0.32.15 | vendored llama.cpp 9d77fa172 | normalized | 4,096 | 102.21 | 101.23 | 5,038.3 | 452 | null | 10,356 | 1,053 | 5.52 | 5.68 | 1,617.6 | 5 | 102.21|102.21|102.17|102.21|102.21 |
llama3.2:3b | null | null | null | null | null | 318.51 | 313.25 | null | 433 | null | 4,343 | null | 3.2 | null | null | 5 | 319.0|318.7|318.2|318.5|317.5 |
llama3.2:3b | null | null | null | null | null | 338.22 | 336.01 | null | 407 | null | 4,339 | null | 2.5 | null | null | 5 | 339.2|338.2|335.0|338.2|339.7 |
qwen2.5:14b | null | null | null | null | null | 99.43 | 98.59 | null | 451 | null | 10,941 | null | 5.6 | null | null | 5 | 99.5|99.5|99.4|99.4|99.3 |
qwen2.5:14b | null | null | null | null | null | 101.38 | 101.29 | null | 423 | null | 10,942 | null | 8.1 | null | null | 5 | 101.3|101.2|101.4|101.4|101.4 |
gpt-oss:20b | null | null | null | null | null | 195.79 | 194.59 | null | null | null | 14,152 | null | 8.6 | null | null | 5 | 195.8|195.7|196.3|196.1|195.7 |
gpt-oss:20b | null | null | null | null | null | 221.96 | 220.53 | null | null | null | 13,037 | null | 19.8 | null | null | 5 | 185.8|242.6|222.0|231.9|175.3 |
# GPU: retail RTX 5080; desktop baseline 1642 MiB GPU memory at bench start (nvidia-smi memory.used | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
# settings: temperature 0, seed 42, n_predict 256 (64 on the long-prefill pass), ctx 4096, full offload, per-iteration prompt nonce, cache_prompt false on llama-server; prefill_long_tps = median of 3 fresh runs after a discarded warmup | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
# prefill_short (~40-token prompt) is dominated by per-request overhead; prefill_long (~420 tokens) is the throughput figure - see the article | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
# third instrument, llama-bench on Llama 3.2 3B: pp512 19589 +/- 2071 t/s, tg256 367.09 +/- 0.69 t/s (raw engine loop, build 10507) | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
# ollama gpt-oss:20b blob refuses to load in upstream llama.cpp: unknown model architecture gptoss (ollama arch tag); its llama.cpp row uses the upstream ggml-org MXFP4 GGUF | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
# gpt-oss MoE decode on upstream llama.cpp is prompt-dependent: fixed prompt 232-242 t/s (median 238.5); varied prompts bimodal (median 235.3) with a ~176 slow path; published 222 median includes a cold first run. Ollama fork: flat 195.7-196.3 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
# 2026-08-20, TechFuelHQ, CC BY 4.0 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
Ollama vs llama.cpp vs LM Studio on One RTX 5080
Controlled runtime comparison of Ollama, llama.cpp and LM Studio on the same RTX 5080 using identical hard-linked, sha256-verified GGUF bytes.
Why this exists
Most runtime comparisons do not control the model file, so you never know whether you measured the runtime or an uncontrolled quantization difference.
Here the same GGUF files are shared between apps via hard links and verified by sha256, so every runtime provably decodes identical bytes and the delta is attributable to the runtime.
Method
- temperature 0, seed 42
- a per-iteration prompt nonce. Without it, warm repetitions hit the prompt cache and report roughly 60k tok/s of "prefill" that measures nothing
- fresh server per model
- wall-clock cross-check on every run
- llama-bench as a third instrument on the 3B
Results
2026-08-20 — llama.cpp b10507 decodes 2-6% faster than Ollama 0.32.1 on dense models.
2026-08-26 — Ollama 0.32.15 decodes 5-11% faster than LM Studio 0.4.21 (351 vs 316 tok/s on Llama 3.2 3B), while LM Studio prefills faster.
The result that surprised me: Ollama gained 3-10% decode on itself between 0.32.1 and 0.32.15 when it updated its vendored engine. That is larger than some of the cross-runtime gaps, so pinning your version matters more than which runtime you are loyal to.
Published caveat
Ollama's gpt-oss blob carries its own gptoss architecture tag, so llama.cpp cannot be pointed at the same bytes there. On the upstream GGUF it decodes 14% faster, but that is a conversion difference and is labelled as one rather than counted as a runtime win.
Negative result retained
A background indexing job was caught depressing 3B decode by ~15% mid-capture. It was suspended and every published cell re-measured in the quiet window. The bad pass stays in the log.
Source and license
Canonical page, full method and change log: https://techfuelhq.com/data/llm-server-compare/
Licensed CC BY 4.0. Attribution: "TechFuelHQ" linking to the canonical page.
Corrections welcome. If a row is wrong it gets fixed against the primary source and the correction is logged.
- Downloads last month
- 12