Mistral’s Timestral Tops TimesFM-3 With 209M Params [2026]

Mistral AI released Timestral, a family of zero-shot time-series forecasting models, claiming its 209-million-parameter flagship outperforms Google's larger TimesFM-3 on the GIFT-Eval, TIME, and fev-bench benchmarks. The Paris-based company announced the release on October 19, 2026, with the Timestral-209M repository dated October 15, 2026 on Hugging Face; the family ships in four sizes — 5 million, 28 million, 123 million, and 209 million parameters — and supports probabilistic outputs, multivariate forecasting, and past and future covariates. Mistral AI's largest Timestral variant is more than two orders of magnitude smaller than the 41 billion active parameters of its own Mistral Large 4 chat model, framing the release as a bet that specialized small architectures can match larger generalist models on forecasting.

Mistral AI has released Timestral, a family of time-series forecasting models that the Paris-based company says beats Google’s TimesFM-3 on three major public benchmarks, despite using a fraction of the parameters. The announcement, reported by MarkTechPost and other outlets on October 19, 2026, positions Timestral as Mistral’s first dedicated push into time-series forecasting, a corner of AI that has quietly become one of the most commercially important battlegrounds in the industry. The pitch is simple: a 209-million-parameter model, according to Mistral AI, outperforms Google’s much larger TimesFM-3 on GIFT-Eval, TIME, and fev-bench, the three benchmarks most commonly cited in forecasting research right now. If that claim holds up under independent testing, it reframes a debate that has mostly been about scale, and turns it into a debate about efficiency. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F Mistral AI Unveils Timestral, a Forecasting Model Family Mistral AI describes Timestral as its first family of time-series foundation models, built for zero-shot forecasting. In the company’s own words, “Today, we’re releasing Timestral, our first family of time-series foundation models,” a line taken directly from the official Mistral AI announcement https://mistral.ai/news/timestral/?ref=upstract.com . The repository for the flagship model, Timestral-209M, is dated October 15, 2026 on Hugging Face, putting the release just four days before this report. Unlike the large language models that made Mistral AI’s name, Timestral is not built for chat or code generation. It is built to take a sequence of numbers, say daily retail sales, hourly power demand, or a stock’s trading volume, and predict what comes next. That is a very different discipline from the transformer-based chat models that dominate headlines, and it draws on a separate lineage of research that includes Google’s TimesFM line, Amazon’s Chronos, and Salesforce’s Moirai. Mistral AI says Timestral supports zero-shot forecasting, probabilistic outputs, multivariate forecasting, and both past and future covariates. In plain terms, that means the model can forecast a new dataset it has never seen without retraining, can express uncertainty rather than a single point estimate, can track multiple related series at once, and can factor in outside variables, like a holiday calendar or a weather forecast, that influence the number being predicted. Inside the Timestral Lineup: 5M to 209M Parameters Timestral ships in four sizes: 5 million, 28 million, 123 million, and 209 million parameters. That range is the headline detail distinguishing Timestral from most frontier AI releases of 2026, where the trend line for language models has pointed toward hundreds of billions or even trillions of parameters. Mistral AI’s own flagship chat model, Mistral Large 4, nicknamed Le Chonk, debuted at 675 billion total parameters https://shattered.io/mistral-large-4-le-chonk-675b-41b-2026/ with 41 billion active at inference. Timestral’s largest variant is smaller than Le Chonk’s active parameter count by more than two orders of magnitude. That gap is the point. Mistral AI’s argument, as described in the announcement, is that time-series forecasting does not need the same parameter budget as natural language generation. A sales curve or a sensor reading has far less entropy than a sentence of English text, and Mistral AI appears to be betting that smaller, specialized architectures can close the gap with larger generalist models on this narrower task. The four-size lineup also opens the door to deployment scenarios that a 200-billion-parameter language model could never touch. A CPU-focused writeup from Lumien AI https://lumienai.com/news/mistral-timestral-time-series-forecasting-model-cpu frames Timestral-209M as “CPU-ready,” meaning it is small enough to run forecasting inference without a dedicated GPU, a detail that matters enormously for companies running thousands of forecasts a day on commodity hardware rather than accelerator clusters. A separate technical walkthrough published by Authon’s engineering blog https://blog.authon.dev/mistral-timestral-209m-zero-shot-forecasting-cpu covers running Timestral-209M locally in Python, underscoring that the model’s small footprint is already drawing developer attention beyond Mistral AI’s own marketing. How Timestral Performs on GIFT-Eval, TIME, and fev-bench Mistral AI evaluated Timestral against three benchmarks that have become the de facto scoreboard for time-series foundation models: GIFT-Eval, TIME, and fev-bench. GIFT-Eval and TIME both score models on MASE, a measure of forecast accuracy relative to a naive baseline, alongside a probabilistic metric that checks how well a model’s uncertainty estimates match reality. Lower MASE scores are better, since they indicate the model’s errors are smaller relative to the baseline it is being compared against. On GIFT-Eval, Mistral AI reports Timestral-209M scoring a MASE of 0.658 and a CRPS of 0.446. On the TIME benchmark, Timestral-209M scored a MASE of 0.626 and a CRPS of 0.522. On fev-bench, a comparison table circulating alongside the release lists a SQL skill score of 50.39 for Timestral-209M. According to Mistral AI, that combination places Timestral-209M first among pretrained, zero-shot models across all three evaluations, using a single checkpoint with no ensembling of multiple models to inflate the score. That last detail, no ensembling, is worth pausing on. In many benchmark leaderboards, teams submit ensembles, multiple models voted or averaged together, that outperform any single model but are impractical to deploy in production. Mistral AI’s claim that Timestral-209M leads with one checkpoint is a claim about real-world usability, not just leaderboard position. The Benchmark Scorecard: Timestral vs TimesFM-3 The most direct comparison in the release is against Google’s TimesFM-3, the newest version of Google Research’s time-series foundation model line. Mistral AI’s published numbers show Timestral-209M edging out TimesFM-3 on both GIFT-Eval and TIME, with TimesFM-3 posting a GIFT-Eval MASE of 0.667 against Timestral’s 0.658, and a TIME MASE of 0.640 against Timestral’s 0.626. On fev-bench, the same comparison table lists TimesFM-3’s SQL skill score at 48.99, versus 50.39 for Timestral-209M. | Benchmark | Metric | Timestral-209M | TimesFM-3 | |---|---|---|---| | GIFT-Eval | MASE lower is better | 0.658 | 0.667 | | GIFT-Eval | CRPS lower is better | 0.446 | 0.456 | | TIME | MASE lower is better | 0.626 | 0.640 | | TIME | CRPS lower is better | 0.522 | 0.536 | | fev-bench | SQL skill score higher is better | 50.39 | 48.99 | | Parameters | Model size | 209 million | Not disclosed in Mistral AI’s release | The margins are not dramatic in absolute terms, a few thousandths of a point on MASE and CRPS, but in a field where benchmark leadership is measured in exactly these increments, they are enough for Mistral AI to claim the top spot. Whether those margins hold up outside Mistral AI’s own test harness is a separate question, and one the company’s release does not resolve on its own. A Contested Crown: Google’s Own TimesFM-3 Claims Top Spot Too Here is where the Timestral story gets more interesting than a simple benchmark win. Google Research’s own announcement for TimesFM-3 makes essentially the mirror-image claim: that TimesFM-3 is the top-ranked pretrained foundation model across GIFT-Eval, fev-bench, and TIME, covering both point forecasts and probabilistic metrics. Google’s blog post describes TimesFM-3 as a model that “enables accurate multivariate time series forecasting in a single forward pass, significantly outperforming other forecasting models across major benchmarks.” Two vendors, two benchmark suites that overlap almost exactly, two opposite conclusions. That is not unusual in AI releases, where companies frequently cherry-pick test conditions, checkpoint versions, or evaluation windows that favor their own model. What it means in practice is that Timestral’s claimed lead over TimesFM-3 should be read as Mistral AI’s own result, not yet an independently adjudicated fact. Neither company’s self-reported numbers should be taken as the final word until third-party researchers reproduce the comparison under matched conditions. This pattern of dueling benchmark claims is becoming familiar across the AI sector. Readers who followed the back-and-forth between DeepSeek’s V4.1-Flash-1017 release and its rivals’ competing benchmark scores https://shattered.io/deepseek-v4-1-flash-1017-hle-score-42-6-2026/ will recognize the shape of this dispute: a headline number, a competing headline number, and a gap that only independent testing can close. The Case for Small Models: Beating Chronos-2 and Falcon-2.0 Mistral AI’s release does not stop at the top-end comparison with TimesFM-3. The company also claims that Timestral-209M outperforms Falcon-2.0, another pretrained forecasting competitor, despite having fewer parameters. And at the small end of the lineup, Mistral AI claims its 28-million-parameter Timestral variant beats Chronos-2, a competing forecasting model that is roughly four times larger, on both GIFT-Eval and TIME. If that smaller-model claim holds up, it is arguably the more interesting result of the two. Beating a much larger rival at the flagship tier is one thing; beating a model four times your size at the entry tier suggests the efficiency gain is not a one-off trick tied to Timestral’s biggest checkpoint, but something closer to an architectural or training advantage that scales down as well as up. Independent benchmarking firms have not yet published a reproduction of either comparison, so this remains a claim to track rather than a settled result. The small-model efficiency argument echoes a theme that has run through several AI releases this year. Nvidia’s push with its Nemotron family made a similar case for punching above a model’s weight class, and readers tracking that story can see the parallel in how Nvidia’s Nemotron 3.5 shipped at 30 billion parameters while aiming at results associated with trillion-parameter systems https://shattered.io/nvidia-nemotron-3-5-30b-1-trillion-param-bet-2026/ . Forecasting is now getting its own version of that argument, with Timestral’s 28-million-parameter model as the test case. Mapping the Time-Series Foundation Model Race Timestral and TimesFM-3 are not the only entries in this field, and understanding where they sit requires a map of the broader landscape. Chronos and its successor Chronos-2 popularized a language-model-style approach to forecasting, tokenizing numeric sequences the way a chatbot tokenizes words, then applying transformer architectures built for text to numbers instead. Moirai, associated with Salesforce’s research group and the open Uni2TS project, took a more general-purpose approach aimed at a wide range of forecasting tasks rather than a single benchmark suite. Nixtla’s TimeGPT occupies a different niche entirely: rather than competing purely on open benchmark scores, it is positioned as a commercial, API-first forecasting service built for production deployment, competing on ease of integration as much as raw accuracy. IBM’s Tiny Time Mixers, often abbreviated TTM, made the case for compact forecasting models well before Timestral did, emphasizing edge deployment and low-latency inference over benchmark-topping scale. Lag-Llama, meanwhile, has carved out a role as a probabilistic, research-oriented model focused on distributional forecasts and uncertainty quantification rather than point accuracy alone. What ties all of these together is a shift in what counts as a competitive edge. A year or two ago, a new forecasting model’s pitch was almost always about accuracy on a single benchmark. Now the pitch increasingly includes multivariate support, covariate handling, probabilistic calibration, zero-shot transfer across unfamiliar datasets, and inference cost, exactly the list of capabilities Mistral AI highlights for Timestral and that Google highlights for TimesFM-3. Where Timestral and Rivals Stack Up The table below lays out how the major players in time-series foundation models currently position themselves, based on each vendor’s own published materials and benchmark claims. Parameter figures for competing models beyond Timestral and TimesFM-3 are not precisely disclosed in the sources reviewed for this report, so they are described qualitatively rather than with a specific number. | Model | Maker | Size range | Positioning | |---|---|---|---| | Timestral | Mistral AI | 5M to 209M parameters | Compact, zero-shot, multivariate, probabilistic forecasting; claims top GIFT-Eval, TIME, fev-bench results | | TimesFM-3 | Google Research | Larger than Timestral-209M, exact figure not published in Mistral AI’s release | Native multivariate forecasting in a single forward pass; Google claims top results on the same three benchmarks | | Chronos-2 | Amazon | Roughly 4x Timestral-28M, per Mistral AI’s comparison | Language-model-style tokenization applied to numeric sequences | | Falcon-2.0 | Technology Innovation Institute | Larger than Timestral-209M, per Mistral AI’s comparison | General pretrained forecasting competitor benchmarked by Mistral AI | | Moirai | Salesforce / Uni2TS | Not specified in sources reviewed | General-purpose zero-shot forecasting baseline widely used in research | | TimeGPT | Nixtla | Not disclosed; API-based | Commercial forecasting API focused on production deployment | | Tiny Time Mixers TTM | IBM | Compact, edge-focused | Low-latency forecasting for edge and resource-constrained deployment | Reading this table alongside the benchmark scorecard above makes the Timestral story clearer. Mistral AI is not claiming to have built the biggest or most capable forecasting model in an absolute sense. It is claiming to have built the most efficient one currently near the top of the three leading public benchmarks, a narrower but commercially sharper claim. Mistral AI’s Funding Surge Behind the Forecasting Push Timestral lands at a moment when Mistral AI’s finances look very different from a year ago. The company, founded in 2023, raised funding in 2025 in a round led by chipmaking equipment giant ASML, and reports at the time valued the company at roughly €11.7 billion. In September 2026, Mistral AI announced a further €3 billion funding round, led by Samsung Electronics alongside the EU-backed Scaleup Europe Fund and PSG Equity, pushing its valuation past €21 billion, or roughly $24 billion. That marks close to an 80% increase in euro-denominated valuation in about a year, and outlets covering the round described it as the largest equity funding round ever raised by a privately owned European technology company. Readers who have followed Mistral AI’s funding trajectory closely will recall the earlier milestone covered when Mistral AI first crossed the $24 billion valuation mark while also opening up an AI safety tool https://shattered.io/mistral-24-billion-valuation-ai-safety-tool-2026/ , a sign that the company has been pairing funding announcements with product launches throughout 2026. Timestral fits that same pattern: a specialized model release timed to keep Mistral AI visible against much larger, better-funded rivals in the US and China. The business logic behind Timestral is straightforward even without inside knowledge of Mistral AI’s roadmap. Forecasting is one of the most directly monetizable applications of AI in finance, retail, logistics, and industrial operations, areas where enterprise customers already pay for software and are comparatively easy to convert into paying AI customers. Mistral AI’s broader open-weight strategy, discussed in coverage of how open-source AI releases from Mistral AI and peers reached an 86% share of notable new model launches https://shattered.io/ai-open-source-bet-86-percent-beam-mistral-2026/ earlier this year, suggests Timestral may follow a similar open-weight distribution model, lowering the barrier for enterprises to test it before committing to paid deployment. From ARIMA to Transformers: A Short History of Forecasting AI Time-series forecasting is one of the oldest applied statistics problems there is, and it predates the current AI boom by decades. For most of the 20th century, the dominant tools were statistical methods like ARIMA and exponential smoothing, which required a trained analyst to fit a model to each individual dataset by hand. Those methods are still used today and remain hard to beat on small, simple datasets, but they do not transfer: a model tuned for one retailer’s sales data cannot be dropped onto another company’s inventory data without being refit from scratch. The shift toward foundation models for forecasting, models pretrained on huge volumes of varied time-series data and then applied to new datasets without retraining, is a much more recent development, largely emerging over the past two to three years. Google’s original TimesFM, Amazon’s Chronos, and Salesforce’s Moirai were among the first wave of serious attempts to bring the zero-shot, transfer-learning paradigm that transformed natural language processing into the forecasting world. TimesFM-3 represents Google’s third iteration of that idea, adding native multivariate forecasting and more flexible covariate handling on top of the original architecture. Mistral AI’s entry into this field with Timestral is notable because the company built its entire public reputation on large language models, not forecasting research. Its flagship products have always been generalist chat and reasoning models, including the Large 4 family covered when conflicting reports surfaced over whether Mistral Large 4’s true parameter count matched the company’s official figures https://shattered.io/mistral-large-4-parameter-count-gap-2026/ . Timestral marks a deliberate branch into a specialized, non-language application, suggesting Mistral AI sees forecasting as a distinct product line rather than a side feature of its general models. Why This Matters for Finance, Retail, and Industrial Forecasting The practical stakes of this release extend well beyond leaderboard bragging rights. Enterprises running demand forecasting, inventory planning, energy load prediction, or financial time-series analysis today typically choose between three options: build and maintain custom statistical models in-house, license a commercial forecasting platform like Nixtla’s TimeGPT, or adopt an open foundation model like Chronos, Moirai, or now Timestral and TimesFM-3 directly into their own infrastructure. A smaller, CPU-capable model changes that calculus in a specific way. Running a 209-million-parameter model, or smaller still, in-house on commodity hardware is dramatically cheaper than running inference for a model that needs a GPU cluster, which matters most to mid-sized companies that cannot justify dedicated accelerator spend for a forecasting workload alone. If Timestral’s benchmark claims hold up under independent scrutiny, it gives those companies a credible reason to consider self-hosting a forecasting model rather than paying for a managed API service. There is also a competitive dimension for Google. TimesFM has been positioned as a flagship research output tied to Google Cloud’s broader AI platform ambitions, and a credible challenger claiming benchmark parity, let alone a narrow lead, puts pressure on Google to either refute the comparison publicly or ship an update that reasserts a clearer gap. Expect Google Research to respond with its own counter-benchmarking in the coming weeks, given the pattern of competitive releases seen throughout 2026’s AI model cycle. What’s Still Unconfirmed About Timestral It is worth being precise about what Mistral AI’s release does and does not establish. The company has not published a commercial price for Timestral, nor has it confirmed what, if any, paid API tier will accompany the open model weights. Inference latency figures, memory requirements for self-hosted deployment, and detailed licensing terms are also not clearly confirmed in the official release reviewed for this report. Separately, a third-party technical blog has described Timestral-209M as capable of running inference in roughly 140 milliseconds on CPU, though that figure comes from an independent writeup rather than Mistral AI’s own materials, and has not been independently reproduced at scale. The headline framing of Timestral as simply “a 209M forecasting model” is also an oversimplification worth correcting. The family spans four distinct sizes, from 5 million parameters up to 209 million, and the smaller variants play just as important a role in Mistral AI’s pitch as the flagship, given the company’s claim that its 28-million-parameter model beats a rival roughly four times its size. Treating Timestral as a single model obscures that the real story is about a family of models aimed at different deployment budgets. Finally, no individual executive’s personal involvement in the Timestral project has been confirmed in the sources reviewed for this report, so attributing the release to any single person at Mistral AI beyond the company itself would be speculation. Readers should treat benchmark numbers from both Mistral AI and Google as vendor-reported until a neutral third party, such as an academic lab or an independent AI evaluation firm, reproduces the comparison. What Comes Next for Time-Series AI Several developments look likely in the weeks and months following this release, based on the pattern of competitive dynamics already visible in the broader AI sector this year. - Independent benchmarking groups and academic labs will likely attempt to reproduce the Timestral-versus-TimesFM-3 comparison on GIFT-Eval, TIME, and fev-bench within the next few months, given how closely the two companies’ claims contradict each other. - Google Research is likely to respond to Timestral with either a public rebuttal of the benchmark comparison or a follow-up TimesFM release aimed at reasserting a clearer performance gap, following the same release-and-counter-release pattern seen elsewhere in AI this year. - Expect other forecasting-focused teams, including the groups behind Chronos and Moirai, to publish updated benchmark numbers of their own now that Timestral has shifted the competitive baseline on GIFT-Eval and TIME. - Mistral AI will likely clarify commercial pricing and hosted API availability for Timestral in the coming weeks, following the same pattern it used after prior open-weight model launches this year. - Enterprise adoption of small, CPU-deployable forecasting models is likely to accelerate through 2027 as companies look for ways to cut AI infrastructure costs, a trend Timestral’s release reinforces rather than originates. Frequently Asked Questions What is Mistral AI’s Timestral? Timestral is Mistral AI’s first family of time-series foundation models, built for zero-shot, probabilistic, and multivariate forecasting. It ships in four sizes: 5 million, 28 million, 123 million, and 209 million parameters. Does Timestral really beat Google’s TimesFM-3? Mistral AI’s own benchmark results show Timestral-209M scoring better than TimesFM-3 on GIFT-Eval, TIME, and fev-bench. However, Google’s own TimesFM-3 announcement claims the opposite, that TimesFM-3 leads across the same benchmarks. Neither claim has been independently verified by a neutral third party as of this report. How many parameters does Timestral-209M have compared to TimesFM-3? Timestral-209M has 209 million parameters. Mistral AI’s release does not disclose an exact parameter count for TimesFM-3 in the materials reviewed for this report, though Mistral AI’s framing suggests TimesFM-3 is substantially larger. Can Timestral run on a CPU instead of a GPU? Third-party coverage describes Timestral-209M as CPU-capable, meaning it can run forecasting inference without a dedicated GPU. That makes it considerably cheaper to deploy than larger forecasting or language models that require accelerator hardware. What benchmarks were used to evaluate Timestral? Mistral AI evaluated Timestral against GIFT-Eval, TIME, and fev-bench, three benchmarks widely used to score time-series foundation models on forecasting accuracy and probabilistic calibration. Is Timestral available to the public? The Timestral-209M model repository is listed on Hugging Face https://huggingface.co/mistralai/Timestral-209M with a release date of October 15, 2026. Mistral AI has not confirmed pricing, licensing terms, or a hosted commercial API in the materials reviewed for this report. How does Timestral compare to Chronos-2 and Falcon-2.0? Mistral AI claims Timestral-209M outperforms Falcon-2.0 despite having fewer parameters, and that its smaller 28-million-parameter Timestral model beats Chronos-2, a rival roughly four times its size, on both GIFT-Eval and TIME. Why is Mistral AI building forecasting models instead of only chat models? Forecasting is one of the most directly monetizable enterprise applications of AI, spanning finance, retail, logistics, and industrial operations. Mistral AI’s move into this space follows a major funding round that valued the company at roughly $24 billion in September 2026, giving it resources to expand beyond general-purpose language models into specialized product lines.