Google TimesFM-3 Review 2026: A 330M Zero-Shot Forecasting Model That Tops Three Leaderboards — But the Weights Are Non-Commercial
Google TimesFM-3 Review 2026: A 330M Zero-Shot Forecasting Model That Tops Three Leaderboards — But the Weights Are Non-Commercial
Verdict first: Google TimesFM-3 (released August 31, 2026) is the first version of Google's time series foundation model built natively for multivariate forecasting — every release through TimesFM-2.5 was strictly univariate. It carries only 330 million parameters, was pretrained on more than 1 trillion time points, accepts multiple targets and known-future covariates zero-shot, and decodes its entire forecast horizon in a single forward pass instead of patch by patch. Google reports it ranks #1 among pretrained foundation models on GIFT-Eval, FEV-Bench and TIME, for both point and probabilistic metrics. Two independent findings matter more than the leaderboards: the zero-shot accuracy and the ~5x speedup over TimesFM-2.5 both hold up under testing, while the marquee multivariate and covariate features produced no measurable benefit in the same tests. And the headline complication is licensing — the code is Apache-2.0, but the 3.0 weights are non-commercial and non-production only. For a forecasting model that most teams would want in production, that is the whole story.
TimesFM-3 — 330M parameters, three #1 rankings among pretrained models, and a non-commercial weights license
What Is TimesFM-3?
TimesFM-3 is the third generation of Google Research's Time Series Foundation Model, announced in a Google Research blog post on August 31, 2026 by research scientists Ayush Jain and Rajat Sen. It is a zero-shot forecaster: you hand it historical series and it predicts the future without any task-specific fine-tuning. The lineage goes back to 2024, when the original TimesFM debuted with an ICML 2024 paper. TimesFM-2.5 (September 2025) cut the model from 500M to 200M parameters and pushed context to roughly 16k — and now sits near the bottom of the same leaderboards it once led, which is a fair measure of how fast this field is moving.
At a glance:
| Attribute | TimesFM-3 |
|---|---|
| Developer | Google Research |
| Announced | August 31, 2026 |
| Parameters | 330 million (one count: 330,710,976) |
| Training data | >1 trillion time points (real + synthetic) |
| Backbone | Decoder-only transformer, patch size 32 time steps |
| Attention | Alternating causal temporal (within series) + full variate (across series) |
| Decoding | Non-autoregressive, Contiguous Patch Masking, single forward pass |
| Output | Point forecast + 9 quantiles (10th–90th percentile) per target per step |
| Context | ~15,360 points (longer inputs truncated to most recent) |
| Code license | Apache-2.0 |
| Weights license | timesfm-non-commercial-license-v1.0 (non-commercial, non-production) |
| Distribution | GitHub google-research/timesfm, Hugging Face google/timesfm-3.0-pytorch |
There is no technical report accompanying this release, which is worth flagging up front — it is one of the reasons the benchmark claims have been harder than usual to scrutinize.
The Architecture: Why This Version Is Different
The backbone is the same decoder-only transformer the earlier versions used, and the tokenization is familiar too: contiguous data points are grouped into patches of 32 time steps, with per-series normalization so that series on wildly different scales can be handled in one pass.
The interesting part is what happens after the patches enter the transformer stack, which Google arranges as a 2D grid traversed by alternating attention:
- Causal temporal attention — horizontal, within a single series. A token can only look at past tokens in its own series, which is what prevents information leakage from the future.
- Full variate attention — vertical, across series at a given time step. A token can look at every other series in the dataset at that point in time, which is how the model learns cross-series correlations. This is the mechanism that makes multivariate forecasting possible: a promotion recorded in one series can inform the forecast for another.
Those two mechanisms alternate across several layers, so every token is repeatedly exposed to both its own history and the company it keeps.
Single-pass decoding
Earlier TimesFM versions decoded patch by patch, which introduced latency and — more importantly — compounding error, since each patch was conditioned on the model's own previous predictions. TimesFM-3 adopts Contiguous Patch Masking to generate the entire forecasting horizon in one forward pass. Masked placeholder tokens are appended for the future horizon; target and past-covariate series are masked there, while past-future covariates stay visible. All masked horizon patches are then filled in simultaneously through the alternating attention layers, with no iterative loop.
Known-future covariates and quantiles
For covariates whose future values are already known — promotions, holidays, weather forecasts — TimesFM-3 uses a lookahead construction in which each token concatenates the current patch with future patches, so upcoming known signals are visible during encoding. And rather than a single number, the model emits 9 quantiles from the 10th to the 90th percentile per target series at every horizon step, which is what makes probabilistic forecasting — and by extension inventory buffers and risk planning — possible without a separate model.
What It Actually Does
Three capabilities are native and zero-shot, meaning no fine-tuning per task:
| Capability | What it means | Example |
|---|---|---|
| Multiple targets | Forecast many related series jointly, each with point and quantile outputs | Sales across several ice-cream brands at once |
| Past covariates | Features known only historically | Past foot traffic into the store |
| Past-future covariates | Future values already known at forecast time | A scheduled promotion; a weather forecast |
Google's own illustration is ice cream sales: instead of extrapolating one sales series in isolation, the model draws on related product sales, foot traffic, a promotion schedule and weather. In the worked example, multivariate mode anticipates roughly a 20% lift on promotion days — a lift the univariate baseline does not see, because a univariate model has no way of knowing a promotion is coming.
That example is a good illustration of the value proposition, and also a good illustration of why the independent findings below matter so much: the benefit depends entirely on whether the covariates carry information the model cannot infer on its own.
Benchmarks: Three #1s, Read Carefully
Google evaluated TimesFM-3 on GIFT-Eval, FEV-Bench and TIME, and describes it as the top-ranked model in both point and probabilistic forecasting metrics among all pretrained foundation models on all three. The comparisons include Chronos-2, the Toto 2.0 family and TimesFM-2.5. Two evaluation entries are reported: in univariate mode, where each series is forecast independently, TimesFM-3 already matches or beats competing models; full multivariate mode then achieves the best average rank.
The specific claims: rank #1 on fev-bench across 100 real-world tasks, rank #1 on the TIME benchmark across 50 domain datasets and 98 evaluation tasks, and rank #1 among foundation models on GIFT-Eval. The published TIME average ranks (lower is better) look like this:
| Model | MASE avg rank | CRPS avg rank |
|---|---|---|
| TimesFM-3 | 4.48 | 4.03 |
| Toto-2.0-2.5B | 5.69 | 5.56 |
| Toto-2.0-313M | 6.15 | 6.15 |
| Toto-2.0-1B | 6.22 | 6.65 |
| Chronos-2 | 7.92 | 8.01 |
| TimesFM-2.5 | 9.86 | 11.03 |
That table is a genuine result, and it is also where the careful reading starts. Three caveats are worth holding onto:
- "Best foundation model" is not "best forecasting system." On the full GIFT-Eval leaderboard, which carries in the region of 126 entries, TimesFM-3 sits around 11th for normalized CRPS. The entries above it are agentic systems, LoRA fine-tunes and ensembles — not single pretrained checkpoints. A Google Cloud AI Research submission called STRIDE, which wraps Chronos-2, reportedly edges TimesFM-3 by about 0.3% on normalized CRPS. So the honest phrasing is: the best single pretrained model, in a leaderboard where assembled systems still win.
- The fev-bench claim was hard to verify. At the time of one analysis, the public fev-bench leaderboard had not yet listed TimesFM-3.
- Average rank hides effect size. An average rank tells you how often a model placed well across tasks; it does not tell you how much better it was on any individual task. Critiques of the release make exactly this point, and ask for independent reproduction on domain data. We address that next.
There is also no technical report for this release, which removes the usual place to check training-data composition, evaluation protocol and per-task numbers.
The Independent Test (The Real Story)
One developer ran a genuinely careful independent evaluation, and the results are more useful than the leaderboards. The setup: two public datasets — daily Wikipedia pageviews (71 articles, January 2023 to September 2026) and hourly city temperatures (24 cities, January 2024 to August 2026) — forecast at horizons of 7 and 28 days and 24 and 168 hours, scored with MASE, sMAPE and weighted quantile loss. Crucially, it used rolling forecast origins — 18 daily, 14 hourly, plus 60 more for a contamination check — totaling roughly 22,000 series-forecasts per model.
Why that rigor matters: the author initially used a single origin, and two of three early conclusions flipped when more origins were added. As he put it, "the one-percent differences from a single origin were just that window's luck."
What held up
- Zero-shot accuracy is real. TimesFM-3 beat seasonal naive by 26–30% on daily data and 14–26% on hourly data, and beat a trained LightGBM by 16–29% — with no training or tuning whatsoever.
- Classical models were unstable by comparison. AutoETS beat naive on the Wikipedia data but was 48% worse on temperature. The foundation model's consistency across domains is arguably its most defensible property.
- Speed is the substantive improvement. Forecasting 71 series 28 days ahead took 1.8 seconds on TimesFM-3 versus 9.2 seconds on TimesFM-2.5 — roughly 5x faster. For context: LightGBM 9.3 seconds including training, AutoETS 18.7 seconds, AutoARIMA 378 seconds per origin. The whole thing ran in 1.3 GB of memory on a laptop RTX 4070. The tester called the speed gain "the real improvement in this version," and that is hard to argue with.
- No sign of contamination. Because the model could have memorized training data, the author fixed the calendar window and varied only the year. On 2026 data — which postdates training — TimesFM-3 did slightly better, by 2.2% on Wikipedia and 1.5% on temperature. That is the right kind of negative result.
What did not hold up
Both headline features did nothing on this data.
- Covariates: supplied with calendar features (day-of-week, a weekend flag), three of four test conditions were statistically meaningless and one was significantly worse. The author's explanation is plausible and important: calendar effects are redundant because the model already learns weekly cycles by itself. He notes the caveat honestly — only calendar covariates were supplied, so genuinely exogenous signals "might have been different."
- Multivariate: the same null result. He re-ran the multivariate grouping using correlation-based clustering specifically to give the feature a fairer shot, and got the same pattern.
That leaves a clean split: the accuracy is real, the speed is real, and the two features Google leads with are unproven outside Google's own evaluations. The accuracy edge over TimesFM-2.5 is also worth sizing — statistically significant, but only 3–5%, with a per-series win rate near 56%.
Where it fails
Two failure modes, both instructive. The model flattens peaks: forecasting about 30°C against an actual 35°C. And it missed a genuine news spike entirely — Google article pageviews rising 16x from 17,000 to 285,000, scoring MASE 3.74. The author's judgment is that this is an unforecastable problem rather than a model weakness, which is fair: no amount of pretraining predicts a story going viral. But if your business is exposed to shocks of that shape, a zero-shot forecaster will not save you.
The tester's closing advice deserves repeating, because it applies to every "I ran it and it was great" claim you will read about any model this year: ask how many forecast origins the result rests on.
The License Problem (The Real Headline)
TimesFM-3's source code is Apache-2.0. The pretrained weights are not. They ship under a new timesfm-non-commercial-license-v1.0, restricted to non-commercial, non-production use. This is a deliberate break from the past: every TimesFM checkpoint through 2.5 shipped under Apache-2.0.
What the license restricts, per the terms as reported:
| Permitted | Not permitted without a separate Google license |
|---|---|
| Testing, evaluation and research not tied to commercial gain | Production deployment of any kind |
| Reproducing published results | Use in commercial decision-making |
| Academic benchmarking | Use in client deliverables |
| — | Use in paid products or services |
| — | Derivative versions built from the model |
| — | Third-party mirroring or redistribution |
Two details make this sharper than a typical "research only" clause. First, the restriction explicitly covers commercial decision-making — which is a wide net, because forecasting is decision-making. If your model's output feeds a replenishment plan, a staffing roster or a budget, you are arguably inside the restricted zone even if no customer ever sees the number. Second, "non-production" is not separately defined in the license text; one detailed analysis found it folded into the definition of "Non-Commercial Purpose," which leaves the boundary less clear than the headline suggests.
The reaction was immediate and not warm. One developer running an independent mirror said the new terms do not permit them to continue hosting it. The New Stack's framing — "you can't use it at work (yet)" — captured the gap between topping three leaderboards and being deployable. The most useful reading of Google's strategy is that this is open weights as a funnel, not a gift: release the checkpoint so researchers and evaluators standardize on it, and route production traffic to BigQuery, where TimesFM-3 is slated to land "in the coming weeks" as the core model behind the AI.FORECAST SQL command, replacing TimesFM-2.5. A managed commercial route with Google's own terms, and no local weights to license.
The structural lesson is worth stating plainly: open-weight licenses are decided per release, not permanently. A model you adopted under Apache-2.0 can ship its next version under a non-commercial license, and your upgrade path closes. Teams building on open weights should keep an exit path — and note that this is a pattern, not an outlier: Salesforce's Moirai 2.0 is also non-commercial (CC BY-NC 4.0), while Amazon's Chronos family and IBM's Granite TimeSeries TTM remain Apache-2.0 and are the obvious beneficiaries when private deployment is the requirement.
How to Run It
Installation is standard, and there are two backends:
pip install timesfm[torch] # PyTorch backend
pip install timesfm[mlx] # native Apple silicon backend, no PyTorchThe PyTorch path builds a ModelConfig (checkpoint path, batch size, device), constructs a forecaster, then calls predict_batch with your arrays, a horizon, and return_quantiles. Results expose .forecast and .quantiles. Multivariate input is a (num_variates, context_length) array, with past-only covariates as (n, context_length) and past-and-future covariates as (n, context_length + horizon). The MLX path mirrors the same interface, so moving between an Apple laptop and a CUDA box is mostly a one-line change.
MLX inference throughput on an M4 Max, 330M model, context 512, horizon 64, fp32 with mx.compile:
| Batch size | Latency | Throughput |
|---|---|---|
| 1 | 11.1 ms | 90 series/sec |
| 8 | 19.7 ms | 406 series/sec |
| 32 | 48.1 ms | 666 series/sec |
MLX and PyTorch agree closely — maximum absolute error of 9.5e-7 for medians and 1.8e-6 for quantiles at horizon 64, with covariate paths matching to 1.7e-6.
Two operational limits to plan around. First, inputs longer than the ~15,360-point global context are truncated to their most recent points — note this is slightly shorter than TimesFM-2.5's roughly 16k, so it is not an upgrade on that axis. Second, horizons longer than one output patch are stitched together from multiple patches, and the repository advises checking longer horizons on their own rather than trusting them as one coherent prediction.
One more line from the repository worth quoting, because it comes from Google: "This open version is not an officially supported Google product." Combine that with the license and the picture is consistent — the open release is positioned for evaluation, not for support contracts.
vs Chronos-2, Toto 2.0 and Moirai 2.0
The time series foundation model field is crowded, and the licensing split runs straight through it — which makes licensing a first-class selection criterion, not an afterthought.
| Model | Developer | License | Notes |
|---|---|---|---|
| TimesFM-3 | Non-commercial (weights) | 330M params, native multivariate, tops all three benchmarks among pretrained models | |
| TimesFM-2.5 | Apache-2.0 | Univariate only; now ranks near the bottom of the same leaderboards | |
| Chronos-2 | Amazon | Apache-2.0 (some sources say MIT) | Multivariate with exogenous variables; a wrapper around it reportedly beats TimesFM-3 by ~0.3% normalized CRPS on GIFT-Eval |
| Toto 2.0 | Datadog | Permissive | Trained on ~2 trillion points — the largest open corpus; family spans 4M to 2.5B parameters |
| Moirai 2.0 | Salesforce | CC BY-NC 4.0 (non-commercial) | Native covariate handling, missing values, mixed frequencies |
| Granite TimeSeries TTM | IBM | Apache-2.0 | Commonly recommended when private deployment and commercial rights are both required |
The practical read: if you need commercial rights and private deployment, Chronos-2, Granite TTM and TimesFM-2.5 are the live options, and TimesFM-3 is not — unless you are going through BigQuery. If you are doing research, benchmarking, or evaluating whether this class of model works for your data at all, TimesFM-3 is currently the strongest single checkpoint available and the license costs you nothing.
The Honest Take
TimesFM-3 does three things genuinely well and one thing badly.
Well: it is small (330M), fast (~5x its predecessor), and accurate enough that a zero-shot model with no tuning beat a tuned LightGBM by 16–29% across two very different domains. That last number is the one to remember, because it is the realistic measure of what "zero-shot" buys you in 2026 — real, sizeable, and not magic.
Badly: the licensing. This is a model whose entire value proposition is production forecasting, released with weights that prohibit production use. Structurally it is defensible — Google wants to sell BigQuery — but it means the release cannot be evaluated on the same axis as Chronos-2 or Granite TTM, because they are playing a different game.
And one thing ambiguously: the multivariate and covariate features, which are the entire reason for the "3" in the name. Google's ice-cream example shows a ~20% promotion lift that univariate misses. The one independent test available found no measurable benefit — but it only supplied calendar covariates, which the model can already infer. So the honest position is: the mechanism is sound and the showcase is convincing, but the feature is unproven on data where it would actually matter. If you have genuinely exogenous signals — promotions, price changes, weather, supply constraints — that is exactly the case worth testing yourself before committing. If your "covariates" are day-of-week and holidays, expect nothing.
The speed alone probably justifies evaluation for most forecasting teams. The license decides whether that evaluation ever becomes a deployment.
Pros & Cons
Pros
- Zero-shot accuracy that survives independent testing — 26–30% better than seasonal naive on daily data, 16–29% better than a tuned LightGBM, with no training
- ~5x faster than TimesFM-2.5 — 1.8 seconds versus 9.2 seconds for 71 series, 28 days ahead
- Runs on a laptop — 1.3 GB of memory on an RTX 4070, plus a native Apple silicon backend
- Native multivariate plus past and known-future covariates, all zero-shot, no per-task fine-tuning
- Probabilistic output by default — 9 quantiles per target per horizon step, no separate model needed
- No compounding error — single forward pass instead of patch-by-patch autoregressive decoding
- No evidence of training-data contamination in independently run year-varying tests
- Consistent across domains where classical baselines were not — AutoETS was 48% worse on temperature than on pageviews
Cons
- Weights are non-commercial and non-production — no commercial decision-making, client deliverables or paid products without a separate Google license
- Restrictions extend to derivative versions and third-party mirrors
- "Non-production" is not separately defined in the license, leaving the boundary unclear
- The multivariate and covariate features showed no measurable benefit in the only independent test available
- Accuracy edge over TimesFM-2.5 is small — 3–5%, with a 56% per-series win rate
- No technical report, so training data, evaluation protocol and per-task numbers cannot be checked
- Benchmarks are Google's own — "best foundation model" still places ~11th on the full GIFT-Eval leaderboard, behind assembled systems
- Context is ~15,360 points — slightly shorter than TimesFM-2.5, and longer inputs are truncated
- Long horizons are stitched from multiple patches and the repository advises checking them individually
- Flattens peaks and misses shocks — forecast 30°C against an actual 35°C, and missed a 16x pageview spike
- Not an officially supported Google product, per the repository's own disclaimer
Who Should Use It (And Who Shouldn't)
| If you are… | Verdict |
|---|---|
| A researcher or data scientist benchmarking forecasting models | Use it. Best single pretrained checkpoint available, free for research, laptop-friendly, and the non-commercial license is irrelevant to your use |
| Evaluating whether a zero-shot forecaster can replace hand-built models on your data | Use it — as an evaluation. The accuracy is real and the speed makes iteration cheap. Just know you are evaluating a capability you may not be able to ship |
| Building a production forecasting pipeline that needs local weights | Don't — not with 3.0 weights. Use TimesFM-2.5, Chronos-2 or Granite TimeSeries TTM instead |
| Already committed to Google Cloud and open to managed services | Watch BigQuery. TimesFM-3 is slated to replace 2.5 as the model behind the AI.FORECAST SQL command — that is the intended commercial path |
| Forecasting series exposed to viral shocks or regime changes | Temper expectations. Peaks get flattened; an unforecastable news spike is unforecastable |
| Doing demand planning with real promotion or pricing signals | Test the multivariate path on your own data first. The mechanism makes sense, but the one independent test found no gain — the difference will come down to whether your covariates carry information the model cannot infer |
Frequently Asked Questions
Can I use Google TimesFM-3 commercially?
Not with the default weights. The source code is Apache-2.0, but the TimesFM 3.0 pretrained weights ship under timesfm-non-commercial-license-v1.0, restricted to non-commercial, non-production use. The license text excludes use in commercial decision-making, client deliverables and paid products or services, and it also covers derivative versions built from the model. Commercial use requires a separate license granted by Google. Weights up to and including TimesFM-2.5 remain Apache-2.0 and are the practical fallback for production deployments, as are Amazon Chronos-2 and IBM Granite TimeSeries TTM.
How many parameters does TimesFM-3 have and what context does it accept?
TimesFM-3 has 330 million parameters — one analysis counted 330,710,976 — and was pretrained on more than 1 trillion real and synthetic time points. It groups data into patches of 32 time steps and operates with a global context of about 15,360 points, slightly shorter than TimesFM-2.5's roughly 16,000. Inputs longer than the global context window are truncated to their most recent points. Forecasts longer than one output patch are stitched together from multiple patches, which the repository advises checking separately.
What is new in TimesFM-3 compared with TimesFM-2.5?
Three things. First, TimesFM-3 is the first version natively pretrained for multivariate forecasting — every prior checkpoint through TimesFM-2.5 was univariate only. Second, it accepts past covariates and past-future covariates zero-shot, with a lookahead token construction that lets the model see known future events. Third, it uses Contiguous Patch Masking to decode the entire forecast horizon in a single forward pass rather than patch by patch, cutting latency and preventing compounding error. Google reports it ranks first among pretrained foundation models on GIFT-Eval, FEV-Bench and TIME.
Do the multivariate and covariate features actually improve accuracy?
One independent tester found no measurable benefit under their conditions. Using calendar covariates such as day-of-week and a weekend flag, three of four test conditions were statistically meaningless and one was significantly worse. Re-running the multivariate grouping with correlation-based clustering to give the feature a fairer shot produced the same pattern. The likely explanation is that calendar effects are redundant because the model already learns weekly cycles on its own. The author notes that only calendar covariates were supplied, so genuinely exogenous signals such as promotions or weather might behave differently. Google's own illustration is an ice-cream promotion schedule, where multivariate mode anticipates a roughly 20% lift on promotion days that the univariate baseline misses.
Is TimesFM-3 actually better than TimesFM-2.5 and how fast is it?
Independent testing found the accuracy edge is real but small — roughly 3 to 5% better than TimesFM-2.5 with a per-series win rate near 56%. The speed gain is the substantive improvement: forecasting 71 series 28 days ahead took 1.8 seconds on TimesFM-3 versus 9.2 seconds on TimesFM-2.5, roughly 5x faster, and it beat a tuned LightGBM by 16 to 29% with no training at all. On Apple silicon, MLX inference on an M4 Max reaches 666 series per second at batch 32.
Is TimesFM-3 free?
The model is free to download and use for non-commercial purposes. There is no API pricing because there is no hosted API — you run the weights yourself, or wait for the BigQuery integration, where Google's managed AI.FORECAST path arrives with its own commercial terms. Notably, no BigQuery pricing for TimesFM-3 had been published at the time of writing.
What are the best alternatives to TimesFM-3?
If you need commercial rights or private deployment: Amazon Chronos-2 (Apache-2.0, multivariate with exogenous variables), IBM Granite TimeSeries TTM (Apache-2.0), or TimesFM-2.5 (Apache-2.0, univariate). If you want the largest open training corpus, Datadog's Toto 2.0 family spans 4M to 2.5B parameters. Salesforce's Moirai 2.0 handles mixed frequencies and missing values, but is CC BY-NC 4.0 — also non-commercial.
Sources
- Google Research — TimesFM-3: A Zero-Shot Foundation Model for Multivariate Forecasting (August 31, 2026)
- GitHub — google-research/timesfm (README, MLX benchmarks, license notes)
- DEV Community — I Tested Google's TimesFM-3: The Zero-Shot Claim Holds, the New Features Don't
- Contiguous Patch Masking — reference implementation in the TimesFM repository
- AI Weekly — Google TimesFM-3 tops forecast benchmarks, non-commercial
- AlphaSignal — Google's TimesFM-3 Tops Three Forecasting Leaderboards With 330M Parameters
- Remio — Google Research Launches TimesFM-3, But Its Best Model Comes With a Commercial Constraint
- Mr. Slash — TimesFM 3.0 and the Non-Commercial License (2026)
- Hugging Face — google/timesfm-3.0-pytorch model card
- DataNorth — Google TimesFM-3 adds multivariate forecasting
- DigitalToday — Google unveils TimesFM-3 for multi-variable time series forecasting
- 至顶网 — 谷歌新预测模型超越所有对手，但暂不能商用
- 渠道网 — 谷歌发布 TimesFM-3：首个原生多变量时序大模型，零样本预测登顶多项基准
Disclosure: benchmark rankings are Google's own as published on August 31, 2026, and had not been fully reproduced independently at the time of writing. The GIFT-Eval full-leaderboard placement, the STRIDE comparison and the license interpretation are drawn from third-party analyses rather than from Google's announcement. Independent test results come from a single third-party evaluation on two public datasets — useful, but not a substitute for testing on your own data. License terms govern actual permitted use; read the license text and, for commercial deployments, consult Google directly. Verify all figures against the official repository and blog post before making an engineering or purchasing decision.
评论
发表评论