Tested and discarded
Collection
Models we tested as a Spanish-language Linux sysadmin agent and did not adopt, with the reason. Full notes on each attempt in the dataset below. • 13 items • Updated
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/harbor/harbor.py", line 171, in _split_generators
raise DataFilesNotFoundError("No task.toml or instruction.md files found")
datasets.exceptions.DataFilesNotFoundError: No task.toml or instruction.md files found
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.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.
Tested and discarded models
Models we tested as a local Linux sysadmin agent used in Spanish (opencode with tool calling, one 24–48 GB GPU, a 262,144-token context) and did not adopt. One page per model with what we ran, the numbers that decided it, and whether the problem was the model, the quantization or the engine.
A model discarded here can be the right choice for other uses; several did very well on tests that do not matter for ours. The numbers come from our own
battery (bateria-llm: tool calls with the real opencode request, agent tasks with real command execution, needle retrieval at long context, reasoning,
games) and are not standard benchmarks.
| Model | Tested | Why we discarded it | Cause | Page |
|---|---|---|---|---|
| google/gemma-4-26B-A4B-it (and 12B) | 2026-10-01, BF16 retest 2026-10-26 | needle 4–5/8 at 64k, 2–3/8 at 116–247k; BF16: 3–5/8 at 241–250k | model (BF16 retest) | gemma-4-26b-a4b.md |
| Qwen/Qwen3.5-9B | 2026-10-02 | reasoning far too long (145–160k tokens for 48 problems); tool calls 9–12/15 | model | qwen3.5-9b.md |
| deepseek-ai/DeepSeek-V4-Flash-0731 | 2026-10-10 | out of memory on long prefills (FreeToken, one 96 GB GPU) | engine | deepseek-v4-flash-0731.md |
| IFM/K2-Horizon-MoVA-36B-A4B | 2026-10-09 | 262k does not fit in 48 GB; KLD 0.057 at W4A16 | memory / quantization | k2-horizon-mova-36b-a4b.md |
| XingChen-AGI/Xing4.0-29B-A4B | 2026-10-11 | agent 4/10, games 1 of 4 run, Chinese in Spanish answers | model | xing4.0-29b-a4b.md |
| JetBrains/Mellum2.1-12B-A2.5B-Thinking | 2026-10-11 | tool calls 0/30 without reasoning, needle 1/8 at 117k | model | mellum2.1-12b-a2.5b-thinking.md |
| Cloudflare/clef-omni | 2026-10-11 | 61/68 on our decision cases against 65/68 for clef and clef-flash | model | clef-omni.md |
| DuoNeural/AdQWENistrator-9B | 2026-10-11 | agent 12/30 against 21–23/30 for NeoHorse and Ornith | model | adqwenistrator-9b.md |
| meta-models/Muse-Glimmer-30B | 2026-10-12 | tool calls 8–18/30; vision 4/15 | model | muse-glimmer-30b.md |
| openbmb/MiniCPM-V-4.7-35B-A3B | 2026-10-13 | behind our 35B-A3B models on vision, tool calls and long context | model | minicpm-v-4.7-35b-a3b.md |
| LiquidAI/LFM2.5-24B-A2B | 2026-10-14 | no vision or MTP; agent in Spanish 9–17/30 (8/10 in English); needle 3/8 at 124k | model | lfm2.5-24b-a2b.md |
| poolside/Laguna-S-2.1 | 2026-10-18 | our EXL3 2.25 bpw was unusable (CJK, tool calls 0/20) | our quantization | laguna-s-2.1.md |
Corrections are welcome in the Community tab, especially when a newer version of a model or an engine changes a result.
- Downloads last month
- 52