JacobPEvans/mlx-benchmarks

MLX Benchmarks

Structured benchmark results for local hardware and hosted-provider language model evaluations. Rows include both local-device runs and provider results; recorded host and accelerator fields may be absent for provider rows. Results are published as immutable Parquet shards by the mlx-benchmarks project.

The default configuration opens data/recent/latest.parquet: the canonical shards published within 30 days of the newest one, newest run first. The publish workflow regenerates it from the canonical shards; it adds no rows of its own. Select history to browse every canonical shard, including older hosted-provider evaluations. Optional result and system fields use nulls when a run did not record them. Historical tag columns remain available individually, the tags_json column retains arbitrary tags, and extra_json retains historical columns outside the stable schema.

Original Parquet shards remain in the repository at their original paths. The normalized shards preserve their rows and add null-valued columns where an original shard did not contain a field.

The public run index records evidence status and run-specific context for canonical shards. A scored shard is eligible for scored comparisons; experimental shards are retained for exploration but are not part of scored comparisons; invalid shards failed the evidence checks described by their index caveats; and recovered shards contain recovered results with their limitations recorded. Invalid shards remain available in history, but are excluded from default and hidden from the viewer.

Model context limits

Max recommended context is the longest documented model limit. YaRN-extended context is shown separately and requires engine support; runtime or memory can impose a smaller limit. Native values come from max_position_embeddings in linked configs or model cards. For conversions and quantizations, the base card supplies limits when the derived card has none.

Model	Native context	YaRN extension	Max rec. context	Context sources
nvidia/Qwen3.8-27B-NVFP4	262,144	Not documented	262,144	Card: up to 262K; config
unsloth/Qwen3.8-27B-GGUF	262,144	1,000,000	1,000,000	card; config
mlx-community/Qwen3.8-27B-8bit	262,144	1,000,000	1,000,000	base card; config
unsloth/Qwen3.8-Flash-Next-GGUF	262,144	1,000,000	1,000,000	card; config
ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF	262,144	1,000,000	1,000,000	Base card; base config
nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF	262,144	Not documented	262,144	card

Quickstart

In the dataset SQL console, start with the measurements available for each task and unit:

SELECT
  suite,
  COALESCE(name, metric_name) AS task,
  COALESCE(metric, metric_metric) AS metric,
  COALESCE(unit, metric_unit) AS unit,
  COUNT(COALESCE(value, metric_value)) AS measured_rows
FROM train
WHERE COALESCE(value, metric_value) IS NOT NULL
GROUP BY
  suite,
  COALESCE(name, metric_name),
  COALESCE(metric, metric_metric),
  COALESCE(unit, metric_unit)
ORDER BY suite, task, metric, unit;

Inspect individual throughput runs with their recorded workload:

SELECT
  timestamp,
  COALESCE(name, metric_name) AS task,
  COALESCE(metric, metric_metric) AS metric,
  model, quantization, chip, engine,
  concurrency AS agents, tag_context_tokens_actual AS context_tokens,
  tag_prompt_tokens AS prompt_tokens, tag_max_gen_toks AS output_budget_tokens,
  tag_truncated_rate,
  COALESCE(value, metric_value) AS value,
  COALESCE(unit, metric_unit) AS unit
FROM train
WHERE suite = 'throughput' AND COALESCE(value, metric_value) IS NOT NULL
ORDER BY timestamp DESC
LIMIT 50;

Keep the unit and workload fields with each result when comparing rows. A null context, prompt, output budget, or completion field means that detail was not recorded for that run; it is not a zero or a value shared with other runs.

Result fields

The normalized default split uses the flat result columns name, metric, value, and unit. In original legacy shards, the equivalent nested result fields may appear as metric_name, metric_metric, metric_value, and metric_unit. When combining downloaded current and legacy shards, coalesce each pair and keep the corresponding unit:

COALESCE(name, metric_name) AS name,
COALESCE(metric, metric_metric) AS metric,
COALESCE(value, metric_value) AS value,
COALESCE(unit, metric_unit) AS unit

Optional system and workload columns include machine label, chip, gpu, engine, quantization, concurrency, and projected tag_* fields such as tag_context_tokens_actual, tag_prompt_tokens, and tag_max_gen_toks. These fields are null when the run did not record or use that dimension. tags_json retains arbitrary tags, and extra_json retains historical fields outside the stable schema.

Detailed campaign envelopes add stable campaign_<group>_<field> columns for hardware, software, model, run, speed, resource, quality, provenance, and cost dimensions. campaign_dimension_null_reasons_json retains the reason codes for explicit unknown or inapplicable values. Historical shards receive nulls in new columns unless their run-index entry supplies a per-shard projection.

Cache dimensions describe the host memory limit, prefix caching, KV and weight offload settings, cache workload shape, GPU KV capacity, cache hit rates, occupancy, eviction delay, preemptions, host RAM peak, and first versus later round TTFT. Memory settings use GiB or MiB as named; hit rates are fractions and KV occupancy is a percentage. Fields that were unavailable or inapplicable are null with a corresponding reason code.

Canonical rows use hardware labels for machine identity. Original shards remain at their original paths as historical source data.