Follow the evolution from desktop graphics cards to data-center accelerators. Explore the hardware behind GPU names such as RTX 4090 and H100, then compare measured AI speeds where the model and workload match.
Explore the evolution of GPUs
From desktop cards and integrated Mac systems to data-center accelerators: explore 220 sourced variants and configurations. Choose a point or browse a device to inspect its evidence. Hardware history is available regardless of whether a compatible AI benchmark exists.
220 devices to explore · 192 plotted · 28 without a verified date
Specifications captured October 2026, placed at sourced announcement or catalog launch dates. Mac and MI300A values describe shared CPU/GPU memory; discrete cards use dedicated GPU memory. These are not launch-day snapshots, guaranteed model capacity or measured AI performance. The year axis stays linear. Horizontal marks preserve quarter/year ranges; points sit at their midpoint.
Scroll the chart horizontally to explore later years, or use Browse hardware below.
NVIDIA: circles · AMD: squares · Intel: triangles · Apple: diamonds. Device names and actual-unit values are also available below. Missing dates/specifications remain unknown, never zero; they stay browsable but unplotted. Logarithmic scaling uses positive values only.
NVIDIA H100 SXM 80GB
Market / scope
Data center · accelerator. Capacity is per physical device/module, not an eight-GPU server.
Memory capacity
80 GB · dedicated GPU memory. Advertised capacity is not guaranteed model-usable memory.
2022-03-22. A manufacturer date is not independent confirmation of first shipping; family announcements need not establish every later configuration.
Single physical GPU/module; power is manufacturer specification, not workload measurement. October is expected initial partner-system rollout, not verified universal availability.
This selected catalog covers reviewed GPU/accelerator variants and Mac system configurations from 2014 onward, with some dates unknown. It is not a complete history, stock list or software compatibility list. Market tags are source-based groupings, not evidence of interchangeable memory or model capability. Data-center specifications do not establish home-system suitability.
From a small computer to a server rack
Explore complete computers, then inspect the assumptions behind larger inference and training scenarios. Prices are US USD, checked 18 October 2026, excluding tax and shipping. Listed hardware, quotes and hypothetical totals keep separate labels.
Configuration 1 of 14: Mac mini M5 Pro ·64 GB
Mac mini M5 Pro ·64 GB
Sourced hardware
Preorder · lead time qualified below
Compact macOS desktop. Exact B&H SKU Z1UD00080 is preorder / coming soon, with no delivery date shown.
Tax / shipping excluded; listing is not stock confirmation.
Memory scope
64 GB unified system memory · systemShared CPU / GPU / OS memory; usable model capacity is workload-dependent.
Storage
1 TB
CPU / GPU
M5 Pro 18-core CPU;20-core Apple GPU
OS / software
macOS; MLX / Metal runtime support must be checked. CUDA-only workflows need another backend.
Power
155 W manufacturer maximum continuous power; not measured inference draw.
Included / upgrades
Power cord included; display / input devices separate. Memory chosen at purchase; external storage expansion.
Other entry configurations
M6: 16 GB / 256 GB, 12 CPU / 12 GPU from $899; M5 Pro: 24 GB / 512 GB, 15 CPU / 16 GPU from $1,699. Official launch starting prices, not prices for this 64 GB SKU.
Limits and additional costs
A standalone desk needs a monitor, keyboard and mouse. Illustrative allowance $300–$700, separate from machine price.
One-year Apple US limited warranty; optional coverage, backup storage and electricity extra.
OS / software: Windows 11 Home; CUDA. Verify current Blackwell-compatible drivers / packages.
Power / cooling
Power: 155 W manufacturer maximum continuous power; not measured inference draw.
Power / Cooling: 1000 W included PSU rating;240 mm liquid CPU cooler. PSU rating is not continuous wall draw.
Storage
1 TB
1 TB
Assumptions
No hypothetical setup assumed
No hypothetical setup assumed
Limits / other costs
A standalone desk needs a monitor, keyboard and mouse. Illustrative allowance $300–$700, separate from machine price. One-year Apple US limited warranty; optional coverage, backup storage and electricity extra.
Illustrative monitor / cable allowance $200–$500; separate from listed tower price. VRAM is fixed on the selected card. Multi-GPU support is not assumed; check regional power / cords, ventilation and support.
Unified system memory, host RAM and dedicated GPU memory are different resources. Capacity alone does not establish model fit, speed or quality. Proprietary model services do not establish downloadable weights or self-hosted hardware requirements. Rack power, cooling, networking, facility support and operation costs require their own evidence and quotes.
Choose a comparable workload
These selectors control apples-to-apples benchmark comparisons. Hardware that fits your budget and supports your software comes first when choosing what to buy.
Workload labels show actual input / returned output tokens, then context capacity.
Decode on the tested systems
2 compatible systems
Server decode rate counts the 255 timed intervals between returned tokens. It excludes the first-token interval and client/network time. Mean ± sample SD across three measured repetitions; one warmup is excluded.
Hardware without compatible measurements is omitted. This slice has no measured Apple result. Estimated token rates, sample rates, aggregate serving throughput and latency are separate measurements and are not mixed into these bars.
What these numbers establish
This first slice covers Qwen3.8-27B on two tested Windows systems: five matched workloads, with prompt processing and decode measured separately. Research reviewed Oct 17, 2026.
These are whole tested systems from the original investigator, Hugo Moreira. Their CPUs, Windows builds, Python and harness revisions differ. The 4090 reports a 495 W GPU limit against a 450 W default. The results do not isolate the graphics card as the cause of a difference.
Thirty physical measured trials support twenty phase summaries across five paired workloads. Review checked source metadata and arithmetic; it did not rerun the experiments. Matching published model fingerprints does not prove which bytes were physically loaded.
There is no quality-equivalence result for the quantizations. Prompt/decode timers are not client time to first token or end-to-end response latency. Three repetitions describe within-condition spread, not independent studies or a confidence interval.
Peak power samples are not integrated energy. No tokens/J, exact host purchase cost, performance-per-dollar or payback date is derived here. Historical prices below are separate contextual observations.
More reviewed whole-system comparisons
Choose a model and comparison method. Each view keeps its own workload, measurement boundary and scale. These results describe complete systems.
Qwen3.8-27B IQ2: code/prose and MTP
Source-bound published study of two complete Windows systems
Same generated code bytes across both systems and MTP states. Five measured requests per system.
Arithmetic mean of five server decode rates: 255 output token intervals / decode time for 256 completion tokens. Prompt processing and the first output interval are excluded. SD is sample SD (n−1).
Complete systems differ in CPU, RAM, OS and other host factors; no isolated GPU effect.
Runtime release, commit and archive hashes match. RTX 4090 has a source preflight executable hash; RTX 4070 Ti has no direct executable hash. Identical executed binaries are not attested.
Model identity and 66/66 target-layer placement are source evidence, not independent per-request loaded-artifact checks. Private launch logs were not acquired.
All 20 code requests have the same output digest; this narrow fixture does not establish general quality.
No client latency, energy or general performance claim. Summaries and phase views add no source requests.
Hugo Moreira, Apache-2.0 repository material. Modified for nathanfennel.com: scalar trials and recomputed means/sample SDs; prompts and generated text omitted.
4,096 context; 1 slot; Q8 K/V; flash attention on; batch 512/128; 2/2 threads; fit off; prompt cache and context shift off
Sampling
temperature 0; top-p 1; top-k/min-p 0; seed 42; thinking/tools/MCP/vision off
MTP
off; draft max/min 0
Repetitions
5 measured requests per system; 1 source warmup excluded per artifact
Mac and consumer source reports
Explore published LLM observations whose counting and timing definitions differ from the comparisons above. Each study keeps its own model, quantization and measurement boundary.
These are qualified source descriptions. They do not establish a hardware ranking, output-quality parity or a cross-study speed ratio.
Qwen3-0.6B · MLX on one reported Mac/session
Reported system: Mac16,9 · source label “Mac M4 Max reference”. Physical GPU bin and effective power mode remain unknown.
Source-reported MLX runtime count/rate on one reported Mac/session. Raw native generation time is not serialized; this is not an independently remeasured token timer.
Reported harness context hint: 672. In the retained source, the MLX adapter inherits a no-op prepareContext; effective MLX context capacity and KV allocation are unknown. This is a source-level interpretation, not executed-binary attestation.
The retained harness calls context preparation only before loading a new model. Warm rows do not establish a fresh allocation. The recorded hint is distinct from the reported native prompt count of 19; recipe arithmetic does not attest exact executed input or tokenization.
Reported model: mlx-community/Qwen3-0.6B-4bit · Q4. Reported runtime: mlx-swift. Model and engine revision stamps are source metadata; loaded weights, tokenizer and binary identities remain unverified.
Two warm source-reported MLX request observations from one reported Mac/session. This is a source-local description, not a hardware or cross-runtime ranking, statistically independent sample, quality score or confidence interval.
The recorded source adapter forwards GenerationInfo token count and generation time; the retained runner forms its native rate from runtime-reported count/time with fallback branches. The exact upstream MLX implementation at the recorded engine revision was unavailable after a stopped HTTP 422 lookup. These are source-reported rates, not independently remeasured token timings.
Raw native generateTime is not serialized. No denominator is reconstructed by dividing count by rate. Total generation-call duration and first-to-last chunk-gap rate use different windows and are excluded from this rate projection.
Model repository revision, Q4 label, engine revision and harness stamp are source-reported metadata, not verified loaded weight/tokenizer/binary/executed-harness hashes. Configured context 672 does not attest actual MLX KV allocation, internal batch, placement or physical GPU core count.
The task label and documented short-chat recipe do not attest the exact executed prompt, chat template or input token vector. Reported prompt count is 19; configured output budget and reported generation count are each 128; reported finish reason is length.
Only 200-character thinking-tag output prefixes are available privately. Complete output parity, emitted token IDs, EOS handling, filtering of reasoning/special tokens and completed-answer quality remain unknown. No output text or prefix is distributed here.
Cold process-first MLX generation is excluded. Warm flags and nominal initial/peak/final thermal labels are source observations; two sub-second requests do not certify thermal stability, sustained behavior or physical repeat independence. Exact power mode and GPU bin remain unknown.
LiteRT records are excluded: their capped count is an approximate chunk tally, their prompt count is 20, and their model/quant/runtime recipe differs. Their erroneous iOS artifact stamp is retained only in private evidence; a retrospective Mac-platform recipe correction is not loaded-binary proof. No cross-arm ratio is supplied.
No energy, per-GPU, hardware-scaling or correct-answer allocation is supported. This factual selection adds zero ledger or documentation credit and does not activate an application comparison.
Copyright (c) 2026 john-rocky and ios-llm-benchmark contributors. Codex for Nathan Fennel, 18 October 2026: newly authored minimal factual field selection and explanatory qualifications; no source execution or remeasurement.
The root license covers software and associated documentation. This exact attributed factual selection does not assert a separate benchmark-database license or rights to models, generated output, upstream code or assets.
Compare native completed-image counts and source-run latency within this fixed image-classification task. These fields stay separate from LLM token rates and from reconstructed energy.
Unverified editorial selection of source-native secondary fields. This presentation is not an official benchmark metric or newly verified score. Result not verified by MLCommons Association.
Published NVIDIA submission artifacts; original official-entry citation is distinct from this site's unverified secondary-metric presentation. Source mean/p90 latency and completed count remain source-run fields, not a relabeling of the rounded official primary score.
2 compatible source systems · original source order
NVIDIA Jetson AGX Orin Developer Kit 64G (MaxQ, TensorRT)
Official entry 3.1-0117; source category available; official table literal: Available Cloud. One reported accelerator per node; 64 GB reported system memory.
Complete submitted Jetson system with shared host/accelerator memory; actual scenario placement is not independently attested.
Reported runtime: Jetpack 5.1.1, TensorRT 8.5.2, CUDA 11.4, cuDNN 8.5.0, DALI 1.19.0.
Completed images in source window
416,879
Source mean latency · ms
1.421479
Source p90 latency · ms
1.635652
Separate accuracy run: 76.096% reported (38,048/50,000). TEST01 separately reports 2,469 matching performance entries. Reported accuracy and the TEST01 consistency report come from separate source runs; neither is an independent replay of all original performance-window predictions.
08-09-2023 18:15:39.581 to 08-09-2023 18:25:39.582. Native printed SPL and LoadGen clock validated against UTC client testing-phase epochs; server timezone is retained, not added twice
Original latency units
1,421,479 ns mean; 1,635,652 ns p90. Millisecond display is a unit conversion only.
NVIDIA Orin NX 16G (MaxQ, TensorRT)
Official entry 3.1-0115; source category available; official table literal: Available Cloud. One reported accelerator per node; 16 GB reported system memory.
Complete submitted Jetson system with shared host/accelerator memory; actual scenario placement is not independently attested.
Reported runtime: Jetpack 5.1.1, TensorRT 8.5.2, CUDA 11.4, cuDNN 8.5.0, DALI 1.19.0.
Completed images in source window
418,234
Source mean latency · ms
1.419952
Source p90 latency · ms
1.418809
Separate accuracy run: 76.096% reported (38,048/50,000). TEST01 separately reports 3,431 matching performance entries. Reported accuracy and the TEST01 consistency report come from separate source runs; neither is an independent replay of all original performance-window predictions.
08-11-2023 22:48:03.931 to 08-11-2023 22:58:03.933. Native printed SPL and LoadGen clock validated against UTC client testing-phase epochs; server timezone is retained, not added twice
Original latency units
1,419,952 ns mean; 1,418,809 ns p90. Millisecond display is a unit conversion only.
SingleStream PerformanceOnly source windows: one image per query and maximum one asynchronous query. Hardware-optimized engines may differ; loaded execution and placement are not independently attested.
Native completed-image count, mean latency and p90 latency come from the selected source performance window. Accuracy and TEST01 consistency evidence concern separate source runs; they do not replay every performance-window prediction.
The source runs share a published classification task and reported-quality basis; hardware-optimized engines may differ. Exact loaded artifacts, preprocessed inputs and complete output replay remain unverified.
A separate source-channel energy reconstruction has two warning-bearing beginning-bracket readings per integral, including one timestamp outside the work window. Physical wiring/calibration and total timing uncertainty remain unresolved. This native count/latency view supports no causal efficiency ranking or per-GPU/token/correct-image allocation. No aggregate uncertainty interval is inferred.
Original system metadata reports available; the official table display literal Available Cloud is retained separately per entry, not a current commercial/orderability claim.
Citation: MLPerf® Inference: Edge v3.1 Closed ResNet50 SingleStream; NVIDIA; closed; MaxQ; source category available; official entries 3.1-0117, 3.1-0115; retrieved 2026-10-18. Source-native secondary fields in an unverified editorial selection.
Newly authored minimal factual source count/latency selection and separate quality qualifications. Full unchanged license and NVIDIA attributions are supplied; no raw source logs, software expression, prediction data, models, images, code, binaries, screenshots or logos are included.
The MLPerf® name and logo are registered and unregistered trademarks of MLCommons Association in the United States and other countries. All rights reserved. Unauthorized use strictly prohibited. See www.mlcommons.org for more information.
Explore an annual electricity-accounting example alongside task measurements. It describes plants operated by one company; it does not identify the electricity supplied to an AI request or data center.
Wisconsin Electric Power Co · 2023
Our sum of EPA plant data for the operator identified by EIA: 21.269 TWh of net generation across 39 plants. This is operated-plant context, not a company-reported portfolio, ownership share, purchases or retail supply.
Generation by source · our calculated shares
Annual net generation; only positive components are plotted. Source component rounding is preserved, without forcing the shares to sum to 100%.
coal47.788%
oil0.015%
gas46.500%
hydro1.680%
biomass0.647%
wind3.228%
solar0.142%
Bar scale: 0–48% of the selected operator's plant generation. All eleven components, including source zeros, remain in the JSON export. Stored decimal tails are not physical measurement precision.
Plants with reported emission rates
Nine plants have compatible native EPA-calculated CO₂ and CO₂e rates. The other thirty have blank emissions fields: missing, not zero. No complete operator-wide intensity can be calculated from this slice.
These are annual operational rates per MWh of net generation, with EPA's biomass and combined heat and power (CHP) adjustments. They are not lifecycle factors. CO₂e uses 100-year AR5 warming potentials without climate-carbon feedbacks.
9 of 9 eligible plants shown · alphabetical order
Concord
EPA plant 7159 · WI (source-reported state)
CO₂ · lb/MWh
1629.302
CO₂e · lb/MWh
1630.805
Native values and source cells
CO₂: 1629.3019999999999 (PLCO2RTA, BC12322). CO₂e: 1630.8050000000001 (PLC2ERTA, BF12322). PLNT23 sheet. Display rounding only; original stored values remain in exports.
Elm Road Generating Station
EPA plant 56068 · WI (source-reported state)
CO₂ · lb/MWh
1933.471
CO₂e · lb/MWh
1948.377
Native values and source cells
CO₂: 1933.471 (PLCO2RTA, BC12344). CO₂e: 1948.377 (PLC2ERTA, BF12344). PLNT23 sheet. Display rounding only; original stored values remain in exports.
Germantown Power Plant
EPA plant 6253 · WI (source-reported state)
CO₂ · lb/MWh
2272.689
CO₂e · lb/MWh
2275.700
Native values and source cells
CO₂: 2272.6889999999999 (PLCO2RTA, BC12364). CO₂e: 2275.6999999999998 (PLC2ERTA, BF12364). PLNT23 sheet. Display rounding only; original stored values remain in exports.
Paris
EPA plant 7270 · WI (source-reported state)
CO₂ · lb/MWh
1606.825
CO₂e · lb/MWh
1608.431
Native values and source cells
CO₂: 1606.825 (PLCO2RTA, BC12427). CO₂e: 1608.431 (PLC2ERTA, BF12427). PLNT23 sheet. Display rounding only; original stored values remain in exports.
Port Washington Generating Station
EPA plant 4040 · WI (source-reported state)
CO₂ · lb/MWh
819.841
CO₂e · lb/MWh
820.666
Native values and source cells
CO₂: 819.84100000000001 (PLCO2RTA, BC12435). CO₂e: 820.66600000000005 (PLC2ERTA, BF12435). PLNT23 sheet. Display rounding only; original stored values remain in exports.
Rothschild Biomass Cogeneration Facility
EPA plant 58124 · WI (source-reported state)
CO₂ · lb/MWh
23.268
CO₂e · lb/MWh
48.599
Source flags: CHP and biomass adjustment. Adjusted operational emissions do not represent lifecycle emissions.
Native values and source cells
CO₂: 23.268000000000001 (PLCO2RTA, BC12451). CO₂e: 48.598999999999997 (PLC2ERTA, BF12451). PLNT23 sheet. Display rounding only; original stored values remain in exports.
South Oak Creek
EPA plant 4041 · WI (source-reported state)
CO₂ · lb/MWh
2419.600
CO₂e · lb/MWh
2436.503
Native values and source cells
CO₂: 2419.6 (PLCO2RTA, BC12462). CO₂e: 2436.5030000000002 (PLC2ERTA, BF12462). PLNT23 sheet. Display rounding only; original stored values remain in exports.
Valley (WEPCO)
EPA plant 4042 · WI (source-reported state)
CO₂ · lb/MWh
787.077
CO₂e · lb/MWh
787.875
Source flags: CHP. Adjusted operational emissions do not represent lifecycle emissions.
Native values and source cells
CO₂: 787.077 (PLCO2RTA, BC12481). CO₂e: 787.875 (PLC2ERTA, BF12481). PLNT23 sheet. Display rounding only; original stored values remain in exports.
Whitewater Cogeneration Facility
EPA plant 55011 · WI (source-reported state)
CO₂ · lb/MWh
942.756
CO₂e · lb/MWh
943.709
Native values and source cells
CO₂: 942.75599999999997 (PLCO2RTA, BC12500). CO₂e: 943.70899999999995 (PLC2ERTA, BF12500). PLNT23 sheet. Display rounding only; original stored values remain in exports.
JSON includes generation context and selected rates; CSV contains only the selected native plant rates. Both retain source revisions, accounting scope, missing coverage and attribution.
United States Environmental Protection Agency (2025), eGRID 2023 revision 2, 12 June 2025. Source: U.S. Energy Information Administration (19 August 2025), Form EIA-923 2023 final revised data. Selection, join, sums, shares and display rounding by Codex for Nathan Fennel, 18 October 2026; not EPA/EIA reported operator aggregates or agency endorsement.
Data period 2023; EPA revision 2025-06-12; EIA re-release 2025-08-19; retrieved 2026-10-18. Operator identity comes from EIA Operator Id, not EPA's transmission/distribution-owner or utility fields.
For facility and request-allocation questions, see the AMICE electricity evidence. No supply contract, physical facility mapping or AI emissions allocation is established here.
Why plant operation, company accounts and ownership differ
Wisconsin Electric Power Co (EIA utility 20847) reported 20,243,340 MWh of net generation and 10,463,826 MWh of wholesale purchases in its observed 2023 Wisconsin EIA-861 account. Total sources and total disposition were each 30,707,166 MWh. This company account uses respondent-owned shares and relevant tolling generation; it is a different scope from our EIA-923-operated-plant context.
That account separately reports 22,918,893 MWh of ultimate-customer sales, 6,983,289 MWh of resale, 46,366 MWh consumed by the respondent without charge, and 758,618 MWh of losses. Missing exchange, wheeling, other-transmission-loss and furnished-without-charge cells stay missing. An arithmetic balance does not prove the missing values are zero or identify the electricity delivered to a particular customer.
Generator ownership and plant operation also differ. EIA-860 lists We Power (owner 49861) with a stored ownership fraction of 1 for Port Washington generator 1CT1; EIA-923 reports plant operator 20847. At Weston RICE generator W11, EIA-860 lists Wisconsin Electric Power Co with fraction 0.5; EIA-923 reports operator 20860, Wisconsin Public Service Corp. These generator fractions do not establish equity-weighted annual generation, retail fuel mix or an emissions factor.
Source: U.S. Energy Information Administration, 2023 Forms EIA-860 and EIA-861 data downloaded October 18, 2026; 2023 Form EIA-923 final revision re-released August 19, 2025, retained in the prior R6 snapshot. Selection, role joins, arithmetic checks and explanations are ours. No agency endorsement. EIA-860/861 source revision dates are not established; retrieval dates do not supply them. The JSON export keeps this company account in a separate context, with native cells, missing values, ownership fractions and source pins; it is not reconciled to the operated-plant sum or used to produce an emissions factor.
Provider source references and accounting boundaries
Provider reports use different accounting boundaries. Annual generation, direct carbon-dioxide intensity and lifecycle greenhouse-gas intensity answer different questions. These source references do not identify electricity supplied to an AI request.
EDF: production and accounting references
Our review keeps consolidated-group and mainland-company reporting scopes separate. The available capture does not prove an electricity-supply contract or a facility's supplied mix.
Source references: EDF. Commentary and scope limitations are our own; no publisher endorsement or data-reuse permission is implied.
Fortum: production-boundary references
Power production and total energy production use different output boundaries. Our review keeps those measures separate rather than treating either as a request-level factor.
Source references: Fortum. This is our newly written accounting note, with no source table, image or provider value reproduced.
Vattenfall: lifecycle-method reference
A technology-specific lifecycle assessment differs from direct operational emissions. The functional unit, geography and exclusions must accompany any eventual reported result; a technology result cannot stand in for an annual provider-wide factor.
Source reference: Vattenfall AB. The methodological distinction is our own explanation; no lifecycle value, chart or report excerpt is reproduced.
Environmental declaration registry references
These external references identify publisher-hosted declarations. No declaration data, validity claim, verification seal or registry logo is reproduced here.
External source references: the International EPD System. No affiliation, endorsement or data license is implied.
Model and price context
A selected history for these two product families and the tested model artifact. Manufacturer starting prices and retail asking snapshots keep their own labels. Neither is a verified purchase or the cost of the tested system. Months remain months; gaps are not interpolated.
TechSpot lowest retail asking snapshot. Exact day, retailer set, taxes and shipping unknown. Source percentage/discount columns have inconsistencies; only reviewed dollar cells are used. No exact tested-board or purchase-price join.
TechSpot lowest retail asking snapshot. Exact day, retailer set, taxes and shipping unknown. Source percentage/discount columns have inconsistencies; only reviewed dollar cells are used. No exact tested-board or purchase-price join.
TechSpot lowest retail asking snapshot. Exact day, retailer set, taxes and shipping unknown. Source percentage/discount columns have inconsistencies; only reviewed dollar cells are used. No exact tested-board or purchase-price join.
TechSpot lowest retail asking snapshot. Exact day, retailer set, taxes and shipping unknown. Source percentage/discount columns have inconsistencies; only reviewed dollar cells are used. No exact tested-board or purchase-price join.
TechSpot lowest retail asking snapshot. Exact day, retailer set, taxes and shipping unknown. Source percentage/discount columns have inconsistencies; only reviewed dollar cells are used. No exact tested-board or purchase-price join.
Published model metadata dates the pinned revision to this day. This is an artifact metadata date, not proof of the model family launch or which weight bytes were loaded in a physical run.
