Datasets:

mlx-community
/

mlx-model-explorer-data

	---
	license: cc-by-4.0
	pretty_name: MLX Model Explorer Data
	language:
	- en
	tags:
	- mlx
	- apple-silicon
	- quantization
	- benchmark
	size_categories:
	- n<1K
	configs:
	- config_name: events
	data_files:
	- split: train
	path:
	- data/events/*.parquet
	- data/incoming/*/.parquet
	default: true
	---

	# MLX Model Explorer Data

	An anonymous record of how people use [MLX Model Explorer](https://huggingface.co/spaces/mlx-community/mlx-model-explorer) to choose an MLX model for their Mac: which model families, sizes, quantizations, memory classes and context lengths they look at, and which models they go on to open, compare or download. It also holds the community reports ("it worked", "too slow") and real MLX benchmark results that people choose to contribute.

	The goal is to answer, with data: what is the MLX community actually trying to run, and how well does it work?

	## Quick start

	```python
	from datasets import load_dataset

	df = load_dataset("mlx-community/mlx-model-explorer-data", "events", split="train").to_pandas()
	df = df[df.suspicious_flags.apply(len) == 0] # clean rows only

	# One row per visit: keep the newest summary of each session
	visits = (df[df.event_type == "session"].sort_values("timestamp")
	.drop_duplicates("session_id", keep="last"))
	visits.explode("quants_searched").quants_searched.value_counts(normalize=True)

	# Community MLX benchmarks: median generation speed per model and chip
	bench = df[df.event_type == "mlx_benchmark"]
	bench.groupby(["selected_model", "chip"]).generation_tps.median()
	```

	## What is recorded

	The dataset records outcomes, not clicks. Individual filter changes and page interactions are never logged.

	\| `event_type` \| One row per \| Contents \|
	\| --- \| --- \| --- \|
	\| `session` \| visit \| The filters the visit ended on, every family and quantization it looked at, the hardware class, the top recommendation shown, and the models it opened, compared or visited on Hugging Face. The page sends it when the visitor first opens a model and again when the tab is hidden, if anything changed. Keep the newest row per `session_id`. \|
	\| `feedback` \| report \| Answer to "Have you tried this model?" and how it went \|
	\| `mlx_benchmark` \| measurement \| A result from `mlx_explorer_bench.py`: speed, memory and optionally perplexity \|
	\| `browser_benchmark` \| test run \| The optional 20-second WebGPU compute test \|

	## Layout and updates

	```
	data/incoming/YYYY/MM/DD/<instance>-<UTC time>-<suffix>.parquet # written by the Space, at most hourly
	data/events/YYYY-MM.parquet # one file per finished month
	eval/wikitext-2-raw-v1-test-v1.txt # pinned text for --quality
	```

	The Space buffers rows and writes a new `incoming` file at most once an hour, or sooner after 2,000 rows. Early each month the previous month is compacted: all of its incoming files are merged into `data/events/YYYY-MM.parquet`, re-sent `session` rows are reduced to the newest one per visit, and the incoming files are deleted, all in one commit. Row values are never edited. The `events` config reads both locations, so recent data is always included. Every file has the same columns in the same order.

	## Schema (`schema_version` 2)

	Every row has every column. Columns that don't apply to an event type are null.

	\| Column \| Type \| Rows \| Description \|
	\| --- \| --- \| --- \| --- \|
	\| `timestamp` \| string \| all \| UTC, second precision, ISO 8601 \|
	\| `schema_version` \| int \| all \| `2` \|
	\| `app_version` \| string \| all \| Space version \|
	\| `event_type` \| string \| all \| See above \|
	\| `session_id` \| string \| all \| Random 16-hex ID per browser tab, gone when the tab closes. Null for script submissions \|
	\| `model_family` \| string \| session, feedback, mlx_benchmark \| Final family filter (`Qwen`, `Gemma`, …). On feedback and benchmarks, the family of `selected_model` \|
	\| `parameter_bucket` \| string \| session, … \| `<3B`, `3-8B`, `8-15B`, `15-35B`, `35-70B`, `70B+`, `MoE` \|
	\| `quantization` \| string \| session, … \| `2-bit` … `8-bit`, `16-bit`, `unknown` \|
	\| `target_context` \| int \| session \| 4096 … 262144 \|
	\| `priority` \| string \| session \| `balanced`, `quality`, `speed`, `memory`, `long_context` \|
	\| `sort` \| string \| session \| `recommended`, `popular`, `recent`, `community`, `all` \|
	\| `result_count` \| int \| session \| Models matching the final filters \|
	\| `families_searched` \| list[string] \| session \| Every family the visit filtered on (max 5) \|
	\| `quants_searched` \| list[string] \| session \| Every quantization the visit filtered on (max 5) \|
	\| `distinct_queries` \| int \| session \| Number of distinct filter combinations the visit looked at \|
	\| `hardware_source` \| string \| session, … \| `confirmed` (user picked memory), `detected` (browser estimate), `none` \|
	\| `hardware_memory_class` \| int \| session, … \| RAM class in GB: 8, 16, 18, 24, 32, 36, 48, 64, 96, 128, 192, 256, 512 \|
	\| `hardware_confirmed` \| bool \| session, … \| Whether the memory class was chosen by the user \|
	\| `webgpu_available` \| bool \| session, browser_benchmark \| WebGPU present in the browser \|
	\| `webgpu_score` \| float \| session, browser_benchmark \| Relative WebGPU compute score (quick check on sessions, 20 s test on browser_benchmark). Not MLX, not tokens/sec \|
	\| `gpu_capability_class` \| string \| session, … \| `high`, `mid`, `entry`, `unknown`, bucketed from the quick score \|
	\| `gpu_vendor`, `gpu_arch` \| string \| session, … \| Generic WebGPU adapter identifiers, e.g. `apple` / `metal-3` \|
	\| `browser_family`, `os_family` \| string \| session, … \| Coarse browser and OS family \|
	\| `cpu_cores` \| int \| session, browser_benchmark \| `navigator.hardwareConcurrency` \|
	\| `top_model` \| string \| session \| Top result under "Best fit" for the final filters \|
	\| `top_model_score`, `top_model_fit` \| float, string \| session \| Its 0 to 100 score and fit label, as shown \|
	\| `engine_version` \| string \| session \| Recommendation engine, e.g. `heuristic-v1` \|
	\| `models_viewed` \| list[string] \| session \| Models whose detail panel was opened (max 10) \|
	\| `models_compared` \| list[string] \| session \| Models added to the comparison (max 5) \|
	\| `models_clicked` \| list[string] \| session \| Models whose Hugging Face page was opened (max 5) \|
	\| `selected_model` \| string \| feedback, mlx_benchmark \| `org/name` Hugging Face repo ID \|
	\| `model_name` \| string \| feedback, mlx_benchmark \| Repo name without the org \|
	\| `quant_bits`, `quant_mixed` \| float, bool \| feedback, mlx_benchmark \| From Hub metadata \|
	\| `pipeline` \| string \| feedback, mlx_benchmark \| Hub pipeline tag \|
	\| `hf_downloads_at_selection`, `hf_likes_at_selection` \| int \| feedback, mlx_benchmark \| Hub counts when the row was written (downloads are Hub's trailing 30 days) \|
	\| `tried` \| string \| feedback \| `yes`, `no`, `planning` \|
	\| `outcome` \| string \| feedback \| `worked` or `problem` \|
	\| `quality_rating` \| string \| feedback \| `excellent`, `good`, `acceptable` \|
	\| `failure_reason` \| string \| feedback \| `too_slow`, `too_much_memory`, `low_quality`, `didnt_run` \|
	\| `reported_ram_gb`, `reported_mac_model` \| int, string \| feedback, mlx_benchmark \| Mac memory and chip (`Apple M1` … `Apple M5 Ultra`, `other`) \|
	\| `reported_tokens_per_second`, `reported_context` \| float, int \| feedback \| Self-reported speed and context \|
	\| `notes` \| string \| feedback \| Optional, ≤ 280 characters; emails, links and phone numbers are replaced before storage \|
	\| `benchmark_type`, `benchmark_version` \| string \| benchmarks \| `webgpu_full` / `mlx_lm`; `webgpu-1` / `mlxbench-2` \|
	\| `benchmark_duration_ms` \| int \| browser_benchmark \| Test duration \|
	\| `prompt_tokens`, `generation_tokens` \| int \| mlx_benchmark \| Benchmark settings \|
	\| `prompt_tps`, `generation_tps`, `ttft_ms` \| float \| mlx_benchmark \| Medians over 3 trials \|
	\| `peak_memory_gb` \| float \| mlx_benchmark \| Peak MLX memory \|
	\| `chip` \| string \| mlx_benchmark \| e.g. `Apple M3 Max` \|
	\| `perplexity`, `perplexity_stderr` \| float \| mlx_benchmark \| On `eval_dataset` (`--quality`); only comparable between quantizations of the same base model \|
	\| `eval_dataset`, `eval_tokens` \| string, int \| mlx_benchmark \| Evaluation text version and tokens scored \|
	\| `mlx_version`, `mlx_lm_version`, `macos_major` \| string, string, int \| mlx_benchmark \| Software versions \|
	\| `suspicious_flags` \| list[string] \| all \| Validation flags; empty means the row passed all checks \|

	### `suspicious_flags`

	\| Flag \| Meaning \|
	\| --- \| --- \|
	\| `unknown_model` / `unknown_model_in_session` \| A model ID not found in the `mlx-community` catalogue at the time \|
	\| `implausible_tps` \| Generation speed above what Apple Silicon memory bandwidth (≤ 800 GB/s) allows for that model, with 3× headroom for speculative decoding \|
	\| `peak_memory_exceeds_ram` \| Peak memory above the reported RAM \|
	\| `incomplete_benchmark` / `incomplete_quality` / `incomplete_feedback` \| Missing the model, the measurement or its version \|
	\| `benchmark_type_mismatch` \| e.g. a browser benchmark labelled as MLX \|
	\| `implausible_webgpu_score` / `implausible_perplexity` \| Values far outside observed ranges \|
	\| `conflicting_feedback` \| Both a success rating and a failure reason \|

	Values outside hard bounds (for example negative speeds, unknown enums or oversized payloads) are rejected and never stored. Flagged rows are kept so filters stay transparent, but the Space excludes them from recommendations and statistics.

	## Privacy

	Never collected: names, emails, accounts, IP addresses, precise location, cookies, user-agent strings, device fingerprints, full GPU description strings, or individual clicks. Timestamps are truncated to seconds. Session IDs are random per tab and not persisted. Rate limiting uses an in-memory salted hash that rotates hourly and is never written anywhere. Browsers sending Global Privacy Control or Do Not Track send nothing unless the user explicitly opts in, and anyone can opt out on the page.

	## Benchmark methodology

	Browser (`webgpu-1`). The quick check runs a 256×256 WGSL matrix multiplication for about 1.2 s. The full test runs matrix multiplications at 256, 512 and 1024 plus a 256 MB buffer copy, about 20 s in total. Each phase is normalized to a fixed reference, and the score is the geometric mean × 1000. It measures WebGPU in the browser, not MLX or LLM inference.

	MLX (`mlxbench-2`). `mlx_explorer_bench.py` follows `mlx_lm.benchmark`: it loads the model with `mlx-lm`, builds a random prompt of `prompt_tokens` tokens (seed 0), disables EOS so exactly `generation_tokens` are produced, does one warm-up run, then reports medians over 3 trials, plus peak memory. If generation speed differs by more than 25% between the fastest and slowest trial (usually memory pressure), speed and TTFT are left empty and only memory and perplexity are reported. Chip and RAM class come from `sysctl`.

	Quality (`--quality`). The script tokenizes `eval/wikitext-2-raw-v1-test-v1.txt` from this repo: the first ~120k characters of the wikitext-2-raw-v1 test split (Salesforce/wikitext, CC BY-SA 3.0, revision `b08601e`), SHA-256 `5dacf0e5…c730e8`. It scores the first 16,384 tokens in non-overlapping 1,024-token windows and reports `exp(mean cross-entropy)`. If the tokenizer has a BOS token, every window starts with it. It is a coarse quality signal and doesn't measure reasoning or instruction following.

	`mlxbench-1` added BOS only at the very start of the text, or not at all for tokenizers that don't add it on encode (Gemma). That inflated perplexity for models with a BOS token, Gemma's into the thousands. Those rows are kept but the Space only uses perplexity from `mlxbench-2`. Speed and memory figures from both versions are comparable. Models without a BOS token (Qwen) give the same perplexity in both.

	The first quality rows were contributed by the Space's maintainer from one M3 Max. They cover every quantization in `mlx-community` that declares Qwen3.5-0.8B, 2B or 4B as its base model: MLX-4bit, OptiQ-4bit and MLX-8bit. Anyone can re-run them with the script to check the numbers.

	## Limitations

	- Self-selected and self-reported. The people who use the Space and submit data are not a random sample of MLX users, and feedback is subjective.
	- Estimated memory classes. Browsers can't read unified memory size. `hardware_source = detected` uses Chrome's rounded-down `deviceMemory` (capped at 32 GB). Prefer `hardware_confirmed = true`.
	- Interest ≠ usage. Visits show what people look at, not what they deploy.
	- Lost summaries. A visit whose browser crashes before it opens a model or hides the tab sends nothing.
	- Heuristic recommendations. `top_model_score` depends on `engine_version`.

	## Versioning

	- `schema_version` changes when columns change meaning or are removed. Version 1 was a per-click log used for the first day after launch. It was replaced by this design and its rows were deleted.
	- `benchmark_version` and `engine_version` identify the exact methodology behind each score.

	## License

	CC BY 4.0. Please cite the dataset and link back to the Space.

	---
	license: cc-by-4.0
	pretty_name: MLX Model Explorer Data
	language:
	- en
	tags:
	- mlx
	- apple-silicon
	- quantization
	- benchmark
	size_categories:
	- n<1K
	configs:
	- config_name: events
	data_files:
	- split: train
	path:
	- data/events/*.parquet
	- data/incoming/*/.parquet
	default: true
	---

	# MLX Model Explorer Data

	An anonymous record of how people use [MLX Model Explorer](https://huggingface.co/spaces/mlx-community/mlx-model-explorer) to choose an MLX model for their Mac: which model families, sizes, quantizations, memory classes and context lengths they look at, and which models they go on to open, compare or download. It also holds the community reports ("it worked", "too slow") and real MLX benchmark results that people choose to contribute.

	The goal is to answer, with data: what is the MLX community actually trying to run, and how well does it work?

	## Quick start

	```python
	from datasets import load_dataset

	df = load_dataset("mlx-community/mlx-model-explorer-data", "events", split="train").to_pandas()
	df = df[df.suspicious_flags.apply(len) == 0] # clean rows only

	# One row per visit: keep the newest summary of each session
	visits = (df[df.event_type == "session"].sort_values("timestamp")
	.drop_duplicates("session_id", keep="last"))
	visits.explode("quants_searched").quants_searched.value_counts(normalize=True)

	# Community MLX benchmarks: median generation speed per model and chip
	bench = df[df.event_type == "mlx_benchmark"]
	bench.groupby(["selected_model", "chip"]).generation_tps.median()
	```

	## What is recorded

	The dataset records outcomes, not clicks. Individual filter changes and page interactions are never logged.

	\| `event_type` \| One row per \| Contents \|
	\| --- \| --- \| --- \|
	\| `session` \| visit \| The filters the visit ended on, every family and quantization it looked at, the hardware class, the top recommendation shown, and the models it opened, compared or visited on Hugging Face. The page sends it when the visitor first opens a model and again when the tab is hidden, if anything changed. Keep the newest row per `session_id`. \|
	\| `feedback` \| report \| Answer to "Have you tried this model?" and how it went \|
	\| `mlx_benchmark` \| measurement \| A result from `mlx_explorer_bench.py`: speed, memory and optionally perplexity \|
	\| `browser_benchmark` \| test run \| The optional 20-second WebGPU compute test \|

	## Layout and updates

	```
	data/incoming/YYYY/MM/DD/<instance>-<UTC time>-<suffix>.parquet # written by the Space, at most hourly
	data/events/YYYY-MM.parquet # one file per finished month
	eval/wikitext-2-raw-v1-test-v1.txt # pinned text for --quality
	```

	The Space buffers rows and writes a new `incoming` file at most once an hour, or sooner after 2,000 rows. Early each month the previous month is compacted: all of its incoming files are merged into `data/events/YYYY-MM.parquet`, re-sent `session` rows are reduced to the newest one per visit, and the incoming files are deleted, all in one commit. Row values are never edited. The `events` config reads both locations, so recent data is always included. Every file has the same columns in the same order.

	## Schema (`schema_version` 2)

	Every row has every column. Columns that don't apply to an event type are null.

	\| Column \| Type \| Rows \| Description \|
	\| --- \| --- \| --- \| --- \|
	\| `timestamp` \| string \| all \| UTC, second precision, ISO 8601 \|
	\| `schema_version` \| int \| all \| `2` \|
	\| `app_version` \| string \| all \| Space version \|
	\| `event_type` \| string \| all \| See above \|
	\| `session_id` \| string \| all \| Random 16-hex ID per browser tab, gone when the tab closes. Null for script submissions \|
	\| `model_family` \| string \| session, feedback, mlx_benchmark \| Final family filter (`Qwen`, `Gemma`, …). On feedback and benchmarks, the family of `selected_model` \|
	\| `parameter_bucket` \| string \| session, … \| `<3B`, `3-8B`, `8-15B`, `15-35B`, `35-70B`, `70B+`, `MoE` \|
	\| `quantization` \| string \| session, … \| `2-bit` … `8-bit`, `16-bit`, `unknown` \|
	\| `target_context` \| int \| session \| 4096 … 262144 \|
	\| `priority` \| string \| session \| `balanced`, `quality`, `speed`, `memory`, `long_context` \|
	\| `sort` \| string \| session \| `recommended`, `popular`, `recent`, `community`, `all` \|
	\| `result_count` \| int \| session \| Models matching the final filters \|
	\| `families_searched` \| list[string] \| session \| Every family the visit filtered on (max 5) \|
	\| `quants_searched` \| list[string] \| session \| Every quantization the visit filtered on (max 5) \|
	\| `distinct_queries` \| int \| session \| Number of distinct filter combinations the visit looked at \|
	\| `hardware_source` \| string \| session, … \| `confirmed` (user picked memory), `detected` (browser estimate), `none` \|
	\| `hardware_memory_class` \| int \| session, … \| RAM class in GB: 8, 16, 18, 24, 32, 36, 48, 64, 96, 128, 192, 256, 512 \|
	\| `hardware_confirmed` \| bool \| session, … \| Whether the memory class was chosen by the user \|
	\| `webgpu_available` \| bool \| session, browser_benchmark \| WebGPU present in the browser \|
	\| `webgpu_score` \| float \| session, browser_benchmark \| Relative WebGPU compute score (quick check on sessions, 20 s test on browser_benchmark). Not MLX, not tokens/sec \|
	\| `gpu_capability_class` \| string \| session, … \| `high`, `mid`, `entry`, `unknown`, bucketed from the quick score \|
	\| `gpu_vendor`, `gpu_arch` \| string \| session, … \| Generic WebGPU adapter identifiers, e.g. `apple` / `metal-3` \|
	\| `browser_family`, `os_family` \| string \| session, … \| Coarse browser and OS family \|
	\| `cpu_cores` \| int \| session, browser_benchmark \| `navigator.hardwareConcurrency` \|
	\| `top_model` \| string \| session \| Top result under "Best fit" for the final filters \|
	\| `top_model_score`, `top_model_fit` \| float, string \| session \| Its 0 to 100 score and fit label, as shown \|
	\| `engine_version` \| string \| session \| Recommendation engine, e.g. `heuristic-v1` \|
	\| `models_viewed` \| list[string] \| session \| Models whose detail panel was opened (max 10) \|
	\| `models_compared` \| list[string] \| session \| Models added to the comparison (max 5) \|
	\| `models_clicked` \| list[string] \| session \| Models whose Hugging Face page was opened (max 5) \|
	\| `selected_model` \| string \| feedback, mlx_benchmark \| `org/name` Hugging Face repo ID \|
	\| `model_name` \| string \| feedback, mlx_benchmark \| Repo name without the org \|
	\| `quant_bits`, `quant_mixed` \| float, bool \| feedback, mlx_benchmark \| From Hub metadata \|
	\| `pipeline` \| string \| feedback, mlx_benchmark \| Hub pipeline tag \|
	\| `hf_downloads_at_selection`, `hf_likes_at_selection` \| int \| feedback, mlx_benchmark \| Hub counts when the row was written (downloads are Hub's trailing 30 days) \|
	\| `tried` \| string \| feedback \| `yes`, `no`, `planning` \|
	\| `outcome` \| string \| feedback \| `worked` or `problem` \|
	\| `quality_rating` \| string \| feedback \| `excellent`, `good`, `acceptable` \|
	\| `failure_reason` \| string \| feedback \| `too_slow`, `too_much_memory`, `low_quality`, `didnt_run` \|
	\| `reported_ram_gb`, `reported_mac_model` \| int, string \| feedback, mlx_benchmark \| Mac memory and chip (`Apple M1` … `Apple M5 Ultra`, `other`) \|
	\| `reported_tokens_per_second`, `reported_context` \| float, int \| feedback \| Self-reported speed and context \|
	\| `notes` \| string \| feedback \| Optional, ≤ 280 characters; emails, links and phone numbers are replaced before storage \|
	\| `benchmark_type`, `benchmark_version` \| string \| benchmarks \| `webgpu_full` / `mlx_lm`; `webgpu-1` / `mlxbench-2` \|
	\| `benchmark_duration_ms` \| int \| browser_benchmark \| Test duration \|
	\| `prompt_tokens`, `generation_tokens` \| int \| mlx_benchmark \| Benchmark settings \|
	\| `prompt_tps`, `generation_tps`, `ttft_ms` \| float \| mlx_benchmark \| Medians over 3 trials \|
	\| `peak_memory_gb` \| float \| mlx_benchmark \| Peak MLX memory \|
	\| `chip` \| string \| mlx_benchmark \| e.g. `Apple M3 Max` \|
	\| `perplexity`, `perplexity_stderr` \| float \| mlx_benchmark \| On `eval_dataset` (`--quality`); only comparable between quantizations of the same base model \|
	\| `eval_dataset`, `eval_tokens` \| string, int \| mlx_benchmark \| Evaluation text version and tokens scored \|
	\| `mlx_version`, `mlx_lm_version`, `macos_major` \| string, string, int \| mlx_benchmark \| Software versions \|
	\| `suspicious_flags` \| list[string] \| all \| Validation flags; empty means the row passed all checks \|

	### `suspicious_flags`

	\| Flag \| Meaning \|
	\| --- \| --- \|
	\| `unknown_model` / `unknown_model_in_session` \| A model ID not found in the `mlx-community` catalogue at the time \|
	\| `implausible_tps` \| Generation speed above what Apple Silicon memory bandwidth (≤ 800 GB/s) allows for that model, with 3× headroom for speculative decoding \|
	\| `peak_memory_exceeds_ram` \| Peak memory above the reported RAM \|
	\| `incomplete_benchmark` / `incomplete_quality` / `incomplete_feedback` \| Missing the model, the measurement or its version \|
	\| `benchmark_type_mismatch` \| e.g. a browser benchmark labelled as MLX \|
	\| `implausible_webgpu_score` / `implausible_perplexity` \| Values far outside observed ranges \|
	\| `conflicting_feedback` \| Both a success rating and a failure reason \|

	Values outside hard bounds (for example negative speeds, unknown enums or oversized payloads) are rejected and never stored. Flagged rows are kept so filters stay transparent, but the Space excludes them from recommendations and statistics.

	## Privacy

	Never collected: names, emails, accounts, IP addresses, precise location, cookies, user-agent strings, device fingerprints, full GPU description strings, or individual clicks. Timestamps are truncated to seconds. Session IDs are random per tab and not persisted. Rate limiting uses an in-memory salted hash that rotates hourly and is never written anywhere. Browsers sending Global Privacy Control or Do Not Track send nothing unless the user explicitly opts in, and anyone can opt out on the page.

	## Benchmark methodology

	Browser (`webgpu-1`). The quick check runs a 256×256 WGSL matrix multiplication for about 1.2 s. The full test runs matrix multiplications at 256, 512 and 1024 plus a 256 MB buffer copy, about 20 s in total. Each phase is normalized to a fixed reference, and the score is the geometric mean × 1000. It measures WebGPU in the browser, not MLX or LLM inference.

	MLX (`mlxbench-2`). `mlx_explorer_bench.py` follows `mlx_lm.benchmark`: it loads the model with `mlx-lm`, builds a random prompt of `prompt_tokens` tokens (seed 0), disables EOS so exactly `generation_tokens` are produced, does one warm-up run, then reports medians over 3 trials, plus peak memory. If generation speed differs by more than 25% between the fastest and slowest trial (usually memory pressure), speed and TTFT are left empty and only memory and perplexity are reported. Chip and RAM class come from `sysctl`.

	Quality (`--quality`). The script tokenizes `eval/wikitext-2-raw-v1-test-v1.txt` from this repo: the first ~120k characters of the wikitext-2-raw-v1 test split (Salesforce/wikitext, CC BY-SA 3.0, revision `b08601e`), SHA-256 `5dacf0e5…c730e8`. It scores the first 16,384 tokens in non-overlapping 1,024-token windows and reports `exp(mean cross-entropy)`. If the tokenizer has a BOS token, every window starts with it. It is a coarse quality signal and doesn't measure reasoning or instruction following.

	`mlxbench-1` added BOS only at the very start of the text, or not at all for tokenizers that don't add it on encode (Gemma). That inflated perplexity for models with a BOS token, Gemma's into the thousands. Those rows are kept but the Space only uses perplexity from `mlxbench-2`. Speed and memory figures from both versions are comparable. Models without a BOS token (Qwen) give the same perplexity in both.

	The first quality rows were contributed by the Space's maintainer from one M3 Max. They cover every quantization in `mlx-community` that declares Qwen3.5-0.8B, 2B or 4B as its base model: MLX-4bit, OptiQ-4bit and MLX-8bit. Anyone can re-run them with the script to check the numbers.

	## Limitations

	- Self-selected and self-reported. The people who use the Space and submit data are not a random sample of MLX users, and feedback is subjective.
	- Estimated memory classes. Browsers can't read unified memory size. `hardware_source = detected` uses Chrome's rounded-down `deviceMemory` (capped at 32 GB). Prefer `hardware_confirmed = true`.
	- Interest ≠ usage. Visits show what people look at, not what they deploy.
	- Lost summaries. A visit whose browser crashes before it opens a model or hides the tab sends nothing.
	- Heuristic recommendations. `top_model_score` depends on `engine_version`.

	## Versioning

	- `schema_version` changes when columns change meaning or are removed. Version 1 was a per-click log used for the first day after launch. It was replaced by this design and its rows were deleted.
	- `benchmark_version` and `engine_version` identify the exact methodology behind each score.

	## License

	CC BY 4.0. Please cite the dataset and link back to the Space.