**Model Summary:** Granite-Speech-4.1-2B-NAR is a non-autoregressive (NAR) speech recognition model that formulates ASR as conditional transcript editing. Instead of decoding tokens one at a time, it edits a CTC hypothesis in a single forward pass using a bidirectional LLM, achieving competitive accuracy with faster inference than autoregressive alternatives. The model is based on the **NLE** (Non-autoregressive LLM-based Editing) architecture described in this paper. For applications where accuracy is the primary concern, consider **granite-speech-4.1-2b**, an autoregressive model from the Granite Speech 4.1 family which achieves higher transcription accuracy at the cost of increased inference latency. Granite-speech-4.1-2b produces punctuated and capitalized transcripts, supports AST and keyword-biased recognition, and includes Japanese. When speaker or word-timing information is needed, consider using **granite-speech-4.1-2b-plus**, which extends the above model with speaker-attributed ASR (speaker labels + word transcripts) and word-level timing information.
We have updated the model weights and configuration. The updated model is more accurate, especially for keyword biasing and for word-level timestamps on long samples. It is a drop-in replacement for the previous version, but we have also updated our usage recommendations to get the best performance. In the tables below, GSP v2610 refers to the updated model and GSP v2604 to the original release. Granite-Speech-4.1-2B-Plus (GSP) has similar capabilities to the Granite-Speech-4.1-2B model. The plus model adds two new community-requested rich transcription features that can be activated with a simple prompt change: speaker-attributed ASR (speaker labels and word transcripts) and word-level timing information. Unlike the base model, the plus model doesn't provide punctuation and capitalization.
> 🎉 **NEW (October 21, 2026): Granite Speech 5.0 TurboCTC is now multilingual!** > 🌍 **7 Languages** (En, Fr, De, Es, Pt, It, Ja) > ❓ **Native punctuation & capitalization** > 🚀 **Better English WER (5.51% vs. 5.78%)** > ⏪ **Previous English-only release: revision 947f59a** **Model Summary:** Granite Speech 5.0 TurboCTC is a compact 470 million parameter multilingual ASR model for English, French, German, Spanish, Portuguese, Italian and Japanese with very high inference speed that is well suited for deployment on laptops, smartphones and other edge devices. The model consists of a conformer acoustic encoder with block self-attention, self-conditioning and temporal downsampling with an output layer corresponding to 16,384 multilingual BPE units. It was trained on approximately 153,000 hours of multilingual audio from public corpora using Connectionist Temporal Classification (CTC) and inference is done non-autoregressively with greedy decoding. The model outputs punctuated and capitalized transcripts. We evaluated granite-speech-5.0-470m-turboctc on standard short-form English ASR benchmarks from the Open ASR leaderboard. Despite adding six languages and native punctuation, the multilingual update improves the average English WER over the original English-only release (5.51% vs. 5.78%):
> **Disclaimer:** This model is intended for research and non-commercial use-only. See Granite Speech 5.0 Turbo CTC if interested in commercial use. **Model Summary:** Granite Speech 5.0 TurboCTC Non-commercial is a compact 470 million parameter English ASR model with very high inference speed that is released for research and noncommercial purposes only. We invite users to refer to granite-speech-5.0-470m-turboctc for other use cases. The model consists of a conformer acoustic encoder with block self-attention, self-conditioning and temporal downsampling with an output layer corresponding to 16,384 BPE units. It was trained on approximately 75,000 hours of English audio from public corpora using Connectionist Temporal Classification (CTC) and inference is done non-autoregressively with greedy decoding.
**Model Summary** Granite-SWASH-2B is a 2B parameter decoder-only dense language model with **per-layer Sliding Window Attention (SWA)** and **learnable attention sinks** (LSE-based). Trained on a mix of open source and proprietary data, it is a broad, general-purpose, English-language base model that serves as an early exploration and small-scale preview for upcoming Granite series model releases. Granite-SWASH-2B is based on a decoder-only dense transformer architecture. Core components of this architecture are: GQA, RoPE, MLP with SwiGLU, RMSNorm, shared input/output embeddings, Sliding Window Attention, and Attention Sinks. **Attention Pattern:** 7 full-attention layers (indices 0, 3, 7, 11, 15, 19, 23) and 17 sliding-window layers (window=128). All layers have learnable per-head attention sinks.
**Model Summary** Granite-SWASH-3B-a600M is a 3B parameter decoder-only sparse language model with **per-layer Sliding Window Attention (SWA)** and **learnable attention sinks** (LSE-based). Trained on a mix of open source and proprietary data, it is a broad, general-purpose, English-language base model that serves as an early exploration and small-scale preview for upcoming Granite series model releases. Granite-SWASH-3B-a600M is based on a decoder-only mixture-of-experts transformer architecture. Core components of this architecture are: GQA, RoPE, MLP with SwiGLU, RMSNorm, MoE with shared experts, shared input/output embeddings, Sliding Window Attention, and Attention Sinks. **Attention Pattern:** 8 full-attention layers (indices 0, 3, 7, 11, 15, 19, 23, 27) and 20 sliding-window layers (window=128). All layers have learnable per-head attention sinks.
**Important**: To use FlowState-r1.1 include the `revision=r1.1` argument when loading the model, as demonstrated in the `Recommended Use` section. If no `revision` argument is provided, the version 1.0 is used. - **FlowState**: We present an SSM-based time series foundation model that can be dynamically adjusted to the specific characteristics of the time series during evaluation. - **Functional Basis Decoder (FBD)**: We propose a novel decoder, as a critical component of FlowState, that utilizes a set of continuous basis functions to make continuous forecasts and allow seamless adjustment to specific input characteristics. - **Flexible temporal adaptation**: FlowState can dynamically adjust the context and target length to the timescale of the provided time series. - **Compact and high-performing**: With fewer than 10M parameters and the ability to forecast multiple consecutive patches in parallel, FlowState delivers state-of-the-art accuracy with exceptional efficiency. This model card contains the model weights for the Granite version of the FlowState model, which is suitable for commercial and enterprise use. A research version of the model (available here) is also provided and is intended for research use only. The Granite and research versions achieve the same MASE performance on GIFT‑Eval, indicating comparable point forecasting accuracy. The research version demonstrates slightly improved CRPS performance, which evaluates the quality of probabilistic forecasts. Despite being **more than 10x smaller** than the next best models, FlowState is **better than all Zero-Shot models** with publicly available source code and model weights on the GIFT-Eval Leaderboard in terms of MASE performance, as of Apr. 28th 2026. The Figure compares GIFT MASE Performance vs. model size for FlowState and the most competitive Zero-Shot Models.
The `PatchTSMixer` is a lightweight and fast multivariate time series forecasting model with state-of-the-art performance on benchmark datasets. In this context, we offer a pre-trained `PatchTSMixer` model encompassing all seven channels of the `ETTh1` dataset. This particular pre-trained model produces a Mean Squared Error (MSE) of 0.37 on the `test` split of the `ETTh1` dataset when forecasting 96 hours into the future with a historical data window of 512 hours. The PatchTSMixer model was proposed in TSMixer: Lightweight MLP-Mixer Model for Multivariate Time Series Forecasting by Vijay Ekambaram, Arindam Jati, Nam Nguyen, Phanwadee Sinthong and Jayant Kalagnanam. PatchTSMixer is a lightweight time-series modeling approach based on the MLP-Mixer architecture. In this HuggingFace implementation, we provide PatchTSMixer’s capabilities to effortlessly facilitate lightweight mixing across patches, channels, and hidden features for effective multivariate time-series modeling. It also supports various attention mechanisms starting from simple gated attention to more complex self-attention blocks that can be customized accordingly. The model can be pretrained and subsequently used for various downstream tasks such as forecasting, classification and regression.
`PatchTST` is a transformer-based model for time series modeling tasks, including forecasting, regression, and classification. This repository contains a pre-trained `PatchTST` model encompassing all seven channels of the `ETTh1` dataset. This particular pre-trained model produces a Mean Squared Error (MSE) of 0.3881 on the `test` split of the `ETTh1` dataset when forecasting 96 hours into the future with a historical data window of 512 hours. The `PatchTST` model was proposed in A Time Series is Worth 64 Words: Long-term Forecasting with Transformers by Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, Jayant Kalagnanam. At a high level the model vectorizes time series into patches of a given size and encodes the resulting sequence of vectors via a Transformer that then outputs the prediction length forecast via an appropriate head.
The architecture incorporates the following changes: - residual blocks in the input and output projections - a quantile head to support probabilistic forecasting - enhanced training strategies incorporating contiguous patch masking, and random masking in the forecast period - trained with reconstruction-loss objective - forecast at inference time is cast as "reconstruction" of the masked forecast period while past context is not masked. If the context period has missing values, both the missing timepoints and forecast timepoints are filled in --- thus providing simultaneous imputation and forecast capability. The particular model here was trained on a diverse set of data (see details below) using significantly longer context length than done previously (8192). Here the context includes both the input context as well as the length of the desired forecast horizon. Shorter input series are pre-pended with the mean of the series values in the context and the padded regions are treated as masked. The d_model is set to 1024, patch_length is 16, and we train the quantile head with 99 quantiles. In total, the model has ~260M parameters, ~250M of which are in the core transformer layers. The model architecture is based on a new implementation of a PatchTST-style architecture, which is available in the IBM TSFM repository. Other implementations of PatchTST can be found here: - Official HuggingFace implementation: Contributed by members of the IBM TSFM team. - Original PatchTST Implementation: Implemented by an IBM Research intern who was part of the IBM TSFM team, and the first author of the original PatchTST paper.
PatchTST-FM-r2, a state-of-the-art zero-shot time series foundation model, represents a continuation of the well-recognized PatchTST model series, building on the original PatchTST and its zero-shot variant PatchTST-FM-r1. PatchTST-FM-r2 brings architectural enhancements as well as an expanded training base on top of its predecessor PatchTST-FM-r1. As of August 31, 2026 Granite-TimeSeries-PatchTST-FM-r2 is the **top performing** zero-shot model released under a permissive, commercial-friendly open-source license on the GIFT-Eval benchmark. Granite-TimeSeries-PatchTST-FM-r2 ranks #2 when considering all zero-shot, replicable models (see below for more details). The architectural changes in r2 include: - Introduction of a convolution layer after the multi-head self-attention (MHSA) layer, sandwiched between two "half-step" FFNs. This block, known as the Conformer block, previously gained popularity in speech recognition. The convolution sublayer tends to capture near-term information between patches thus enabling the MHSA to focus on long-term relationships. - Plain-transformer blocks are replaced by Conformer blocks with alternating convolution kernel sizes of 3 and 5, in a repeating pattern {5, 5, 3, 3} - Overlapping patches of size 16 with stride of 8, combined with a Hamming window loss weighting at training time, and an overlap-and-add forecasting approach at inference time - The overall number of blocks increased to 30 (from 20) to absorb expanded training data - A pre-head layer norm is applied for enhanced training stability - Training is expanded by synthetic data generated using the CauKer method, as used in the training of the Granite-FlowState model As with PatchTST-FM-r1, the r2 model was trained on a diverse set of data (see details below) using context length of 8192. The model hidden dimension is set to 1024, patch_length is 16, and the quantile head spans 99 quantiles. In total, the model has ~385M parameters.
**TSPulse** models are ultra-compact pre-trained models for time-series data, featuring just **1M parameters** and supporting **GPU-free inference**. Designed for versatility, TSPulse excels across a wide range of tasks including **classification**, **anomaly detection (AD)**, **imputation**, and **similarity search**. At the architecture level, TSPulse introduces a novel dual-space masked reconstruction strategy that learns jointly from both time and frequency domains —- capturing complementary patterns in a unified embedding space. This is coupled with a dual-embedding disentanglement mechanism that produces both fine-grained embeddings for detailed analysis and semantic embeddings for broader understanding. These semantic embeddings are inherently robust to variations in time, magnitude, and noise, making them ideal for time-series retrieval. At the task level, TSPulse integrates several innovations: - **TSLens**: A fine-tuning module for task-aware feature extraction. - **Multi-head triangulation**: Fuses outputs from multiple prediction streams to enhance anomaly detection robustness. - **Hybrid masking**: Reduces bias during pre-training, improving zero-shot imputation.
TinyTimeMixers (TTMs) are compact pre-trained models for Multivariate Time-Series Forecasting, open-sourced by IBM Research. **With less than 1 Million parameters, TTM (accepted in NeurIPS 24) introduces the notion of the first-ever “tiny” pre-trained models for Time-Series Forecasting.** TTM outperforms several popular benchmarks demanding billions of parameters in zero-shot and few-shot forecasting. TTMs are lightweight forecasters, pre-trained on publicly available time series data with various augmentations. TTM provides state-of-the-art zero-shot forecasts and can easily be fine-tuned for multi-variate forecasts with just 5% of the training data to be competitive. Refer to our paper for more details. **The current open-source version supports point forecasting use-cases specifically ranging from minutely to hourly resolutions (Ex. 10 min, 15 min, 1 hour.).**
Adjust provider, category or input/output filters. New models are added continuously.
COST EXPLORER
Great models. Clear costs.
Cost analysis
Adjust monthly token usage to compare estimated API costs.
1 M = 1 million tokens · USD Compares base text-token rates within the same channel
Uses collected, comparable USD-per-million-token rates. Snapshots exclude cache discounts, tiers, extra reasoning output and taxes. Check current source prices.
MODEL HANDBOOK
Understand before you choose
Selection guide & FAQ
How should API prices be compared?+
Separate input, output and cache rates, then estimate the same workload. Image, audio and video rates may use items, duration or resolution rather than tokens.
Do open weights mean free commercial use?+
Not necessarily. Check each version’s license, commercial restrictions and policies. Self-hosting also incurs compute and maintenance costs.
Is the top-ranked model the best for me?+
Rankings depend on tasks, versions and methods. Compare accuracy, latency, reliability and costs on your own workload.
