TinyTim V1: Divergent LM and Time Mixers

Updated 4 July 2026

TinyTim V1 is a context-dependent term that identifies both a divergent language model and a compact multivariate time series forecasting system.
The language model version is fine-tuned on Finnegans Wake, generating high-variance, lexically inventive outputs through causal LM training.
The forecasting variant, known as Tiny Time Mixers, employs adaptive patching and transfer learning to achieve 12–38% performance gains over state-of-the-art methods.

Searching arXiv for "TinyTim V1" and closely related papers to ground the article. TinyTim V1 is used in recent arXiv literature to denote two technically unrelated systems: a specialized LLM for divergent generation, obtained by fine-tuning TinyLlama-1.1B-Chat-v1.0 on James Joyce’s Finnegans Wake, and, in another paper’s framing, Tiny Time Mixers (TTM), a family of compact pre-trained models for zero-shot and few-shot multivariate time-series forecasting (Agostino, 15 Aug 2025, Ekambaram et al., 2024). This suggests that the name is context-dependent rather than a stable identifier for a single architecture or research program.

1. Terminological scope

The language-model usage is explicit in "TinyTim: A Family of LLMs for Divergent Generation," where TinyTim V1 is the resulting model after fine-tuning on the complete Finnegans Wake corpus. Its stated objective is not helpfulness or semantic stability, but divergent generation: unusual, high-variance, linguistically inventive, and often semantically loose output (Agostino, 15 Aug 2025).

A separate usage appears in "Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series," whose details identify TTMs as also being called "TinyTim V1 / TTM" in that paper’s framing. There the term refers to a lightweight, time-series-native forecasting architecture based on TSMixer, designed for transfer across heterogeneous multivariate datasets (Ekambaram et al., 2024).

Because these usages differ in domain, objective, training data, and evaluation protocol, the term requires immediate disambiguation in scholarly citation. A plausible implication is that references to TinyTim V1 are incomplete unless paired with the relevant arXiv identifier.

2. Divergent-generation LLM

In the language-model literature, TinyTim V1 is built by fine-tuning TinyLlama-1.1B-Chat-v1.0 on the full text of Finnegans Wake—about 1.5 MB of text—segmented into 100-word chunks and trained with a standard causal language modeling objective (Agostino, 15 Aug 2025). The corpus choice is central to the model’s design: Finnegans Wake is treated as an "anti-parsimonious" text whose multilingual, punning, allusive, and structurally unconventional character is intended to pull the model away from ordinary convergent response patterns.

The paper contrasts this design with baseline systems described as convergent: qwen3:0.6b, llama3.2, and gpt-5-mini. In that comparison, baseline models are framed as optimized for consistency and semantic alignment, whereas TinyTim V1 is framed as lexically inventive, high variance, semantically less coherent, and less retrieval-like. The paper summarizes the contrast as follows: the baselines are sophisticated retrievers; TinyTim is a lexical inventor (Agostino, 15 Aug 2025).

The notion of divergent generation is defined operationally rather than metaphorically. It prioritizes novelty over conventionality, associative leaps over direct topical relevance, unusual vocabulary and phrasing over standard fluency, wide output variability over tight response distribution, and semantically loose but creatively suggestive text. The model is therefore not presented as a general-purpose assistant. The paper also notes that TinyTim V1 is not instruction-tuned, while mentioning plans for a future instruction-tuned version that could both "generate text like Joyce" and "answer questions and be helpful" (Agostino, 15 Aug 2025).

3. Quantitative profile of the language-model variant

The evaluation of TinyTim V1 uses 10 creative prompts and a battery of syntactic, lexical, and semantic measures, including Unique Word Ratio (UWR), Average Word Length, Token Diversity via Shannon entropy, Sentence Complexity, Semantic Similarity to the prompt using sentence embeddings, Readability via Flesch-Kincaid Grade Level, and Sentiment via VADER compound score (Agostino, 15 Aug 2025). The study reports 2400 total generated samples, of which 1013 valid samples remained after filtering malformed or single-word outputs. Final sample sizes are reported as TinyTim n = 714, gpt-5-mini n = 99, llama3.2 n = 100, and qwen3:0.6b n = 1100.

Group differences across all metrics were tested with a Kruskal–Wallis test with reported significance $p < .0001$ , followed by pairwise Mann–Whitney U tests with Bonferroni correction. The paper states that TinyTim differed significantly from all three baselines on every metric (Agostino, 15 Aug 2025).

The strongest reported evidence concerns lexical richness.

Measure	TinyTim	Comparator
Hapax legomena ratio	0.643	`gpt-5-mini`: 0.413
Yule’s K	208	`gpt-5-mini`: 47

The paper states that TinyTim’s hapax ratio is over 50% higher than gpt-5-mini, and that its Yule’s K is over four times greater (Agostino, 15 Aug 2025). These results are interpreted as evidence that the model is generating rare, novel, and sometimes invented forms from a constrained learned lexicon.

Distributionally, baseline models are described as having tight, narrow, predictable output distributions, whereas TinyTim has wide, skewed distributions with long tails. Sentence Complexity is reported to range from single-word fragments to sentences over 200 words long. A central trade-off is captured by the reported strong negative correlation between Token Diversity and Unique Word Ratio, $r = -0.807$ , with baselines occupying a high-diversity, low-uniqueness cluster and TinyTim occupying a high-uniqueness, low-diversity region (Agostino, 15 Aug 2025). The paper also states that TinyTim has low semantic coherence relative to the prompt and baselines, reflected in lower prompt semantic similarity and more fragmented or associative outputs.

4. Creativity framework, system role, and limitations

The language-model paper situates TinyTim V1 within theories of divergent vs. convergent thinking, dual-process creativity, and the interaction of the brain’s Default Mode Network (DMN) and Executive Control Network (ECN) (Agostino, 15 Aug 2025). Within that framing, standard LLMs are described as approximating convergent cognition, whereas TinyTim is designed as an artificial substrate for associative ideation. It is not optimized to return the best answer, but to emit surprising raw material.

This interpretation supports the paper’s architectural claim that TinyTim can function as a “divergent knowledge source” inside larger creative systems. The proposed division of labor is explicit: TinyTim produces unusual, high-variance, low-coherence material; another agent—human or convergent LLM—interprets, filters, and integrates it. The paper associates this role with multi-agent creative architectures, automated discovery systems, concept exploration workflows, serendipity engines, and human-AI co-creation systems (Agostino, 15 Aug 2025).

The same paper also states several limitations. These include the narrow training domain of only Finnegans Wake, low semantic coherence, the fact that the model is not instruction-tuned, the small model scale of a 1.1B-parameter backbone, the use of only 10 creative prompts for evaluation, and the absence of a human creativity gold standard. The model had been available on Hugging Face for over a year and had been downloaded more than 750 times at the time of reporting (Agostino, 15 Aug 2025). A plausible implication is that TinyTim V1 is best interpreted as a stylized generator specialized for exploratory ideation rather than as a general benchmark for language-model capability.

5. TinyTim V1 / TTM in multivariate time-series forecasting

In the time-series literature, TinyTim V1 refers to Tiny Time Mixers (TTMs), described as a family of small, fast, pre-trained forecasting models for multivariate time series (Ekambaram et al., 2024). Their central premise is that transfer-learning gains in forecasting need not rely on large LLM-based systems. Instead, the model is built around a lightweight TSMixer backbone and trained exclusively on public time-series datasets.

The forecasting task is formulated with context input $\mathbf{X} \in \mathbb{R}^{c \times sl}$ , target $\mathbf{Y} \in \mathbb{R}^{c' \times fl}$ , and prediction $\hat{\mathbf{Y}}$ , optimized by mean squared error $\mathcal{L} = \left\| \mathbf{Y} - \hat{\mathbf{Y}} \right\|_2^2$ (Ekambaram et al., 2024). Before the backbone, each channel is normalized instance-wise to zero mean and unit variance, then split into non-overlapping patches. The patched input $\mathbf{X}_p \in \mathbb{R}^{c \times n \times pl}$ reduces computation while preserving local semantics.

The most distinctive architectural mechanism is adaptive patching across layers. With $L$ levels and $M$ TTM blocks per level, the model reshapes the patch dimension so that deeper stages observe different granularities, using $K_i = 2^{(L-i)}$ . This is paired with downsampling-based data augmentation, introduced to create additional lower-resolution corpora from high-frequency datasets, and with optional resolution prefix tuning, which embeds the sampling rate as a conditioning prefix. The paper reports that downsampling yields about 30% improvement in the zero-shot setting, while resolution prefix tuning gives about 8% improvement for shorter context lengths and little to no gain for long context windows (Ekambaram et al., 2024).

TTM also adopts a multi-level strategy for transfer. During pre-training, all channels are treated independently and the model is trained in a univariate, channel-independent way. During fine-tuning, the backbone is frozen, and the decoder can operate either channel-independently or with channel mixing for multivariate target datasets. To support known future covariates, the model includes an exogenous mixer block, which replaces exogenous predictions with true future exogenous values, forms overlapping windows, and applies a TSMixer block with channel mixing enabled (Ekambaram et al., 2024).

TTM is reported as a tiny model with $r = -0.807$ 0 million parameters, pre-trained on about 244 million samples from public Monash datasets. Separate pre-trained models are built for $r = -0.807$ 1 configurations $r = -0.807$ 2, $r = -0.807$ 3, $r = -0.807$ 4, $r = -0.807$ 5, and $r = -0.807$ 6, and each model reportedly trains in only 4–8 hours on 6 A100 GPUs (Ekambaram et al., 2024).

Its headline empirical claims are transfer-oriented. On the common benchmark set of ETT, Weather, Electricity, and Traffic, TTM achieves 12–38% improvement in 5% few-shot settings over SOTA and 4–45% improvement in 10% few-shot settings over SOTA. It is also reported to outperform GPT4TS and to exceed LLMTime by about 29% in the cited zero-shot comparison. Against self-supervised pre-training methods such as SimMTM, Ti-MAE, TST, LaST, TF-C, CoST, and TS2Vec, the reported gains are roughly 17–43% (Ekambaram et al., 2024).

Efficiency is a central part of the model’s identity. Relative to GPT4TS, TTM is reported to use 14× fewer learnable fine-tuning parameters, 106× fewer total parameters, 65× faster fine-tuning, 54× faster inference, and 27× lower memory usage. The paper further emphasizes that TTMs are lightweight enough to run in CPU-heavy or CPU-only settings (Ekambaram et al., 2024).

A later mouse-V1 decoding paper introduces Dual-Tower Image–Neural Alignment (DINA) and describes it as TinyTim V1’s interpretive successor in its own framing, but DINA is a distinct interpretable contrastive framework for aligning image feature maps and neural feature maps in mouse primary visual cortex rather than a continuation of the forecasting or language-generation systems (Wang et al., 5 May 2026). This further reinforces the need for disambiguation: "TinyTim V1" does not refer to a single canonical artifact, but to separate domain-specific constructs whose meanings must be recovered from citation context.