Pump.fun: Solana Memecoin Platform

• Pump.fun is a decentralized platform on Solana that transforms internet memes and community interactions into standardized on-chain tokens using a simple fair-launch mechanism.
• The platform employs innovative bonding curves and graduation rules to manage liquidity and mitigate traditional rug-pull risks, ensuring transactional fairness.
• Pump.fun’s standardized design and rich multimodal datasets provide valuable empirical insights into token viability, market dynamics, and community-driven finance.

Pump.fun is a Solana-based memecoin launchpad and trading platform that research literature treats as a paradigmatic interface through which internet memes, community interaction, and speculative finance are converted into standardized on-chain assets. Across studies from 2024 to 2026, it is described both as a decentralized platform that simplifies token creation through a simple web interface and as a market-design environment whose bonding-curve mechanics, graduation rules, and comment-centric user interface make it unusually tractable for empirical analysis of Web3 memecoins (Long et al., 2024, Marino et al., 16 Feb 2026, Mancino, 4 Dec 2025).

1. Definition and position within the Solana ecosystem

Pump.fun is described as a decentralized platform on Solana with a simple web UI that allows a user to create a token by specifying basic fields such as name, symbol, and total supply, while also uploading a logo and description. The platform handles on-chain token creation, so no smart-contract coding or deep blockchain knowledge is required. Once created, each token has a dedicated page that co-locates the logo image, textual description, user comments, and financial statistics, which makes the platform simultaneously a launch mechanism, a social interface, and a market data surface (Long et al., 2024).

The literature emphasizes Pump.fun’s scale. As of September 2024, it had enabled over 1 million memecoin launches and generated \$100M+ in revenue, while serving as a dominant source of new tokens on Solana. In Q4 2024, one study reports a peak of 69,046 token mints in a single day, equal to 71.1% of all tokens minted on Solana during that period. The same study reports that Pump.fun’s own DEX program accounted for 15% to 25% of total DEX activity, processing between 2 and 4 million transactions each day, while transactions involving Pump.fun-created tokens constituted between 40% and 67.4% of daily DEX transactions. Daily active users rose from around 60,000 in early October 2024 to peaks of 260,000 in November and 224,000 in December, reinforcing the platform’s characterization as a retail-oriented engine of memecoin activity (Mancino, 4 Dec 2025).

2. Launch mechanics, bonding curves, and graduation

The platform’s operational logic is described in two closely related ways. In the Q4 2024 Solana trends study, Pump.fun is presented as enforcing a “fair-launch” mechanism with no presale or team allocation: users buy directly on a bonding curve and may sell at any time. In that description, tokens that successfully complete the bonding curve and reach a \$100k market capitalization can “graduate” to Raydium, at which point Pump.fun deposits \$17k of liquidity into Raydium and permanently burns that liquidity. The same study treats this mechanism as an attempt to reduce classic rug-pull modes associated with insider allocations, while still allowing fully speculative secondary trading after graduation (Mancino, 4 Dec 2025).

A later market-design analysis formalizes the mechanism as a constant-product market maker with two phases: a virtual AMM during launch and a real AMM after graduation. In both phases, reserves satisfy

$x\,y = k, \qquad P = \frac{x}{y}.$

At creation, the protocol mints a total supply of

$y_0^{\rm mint} = 10^9 \text{ tokens},$

split into

$y_0^{\rm virt,mint} = 0.7931 \times 10^9$

for the virtual bonding curve and

$y_0^{\rm pswap,mint} = 0.2069 \times 10^9$

for later listing. The virtual pool is initialized with synthetic SOL

$x_0^{\rm synt} = 30 \text{ SOL}$

and synthetic tokens

$y_0^{\rm synt} = 0.2799 \times 10^9.$

Trading fees total 1.25% of exchanged SOL, split into 0.3% to the creator and 0.95% to the platform. In that study, graduation occurs when total virtual SOL reaches 115 SOL, after which 85 SOL of real liquidity seed PumpSwap and trading migrates to a standard on-chain AMM (Marino et al., 16 Feb 2026).

The dataset papers use Raydium listing as the relevant success event, but they do not describe the threshold identically. One dataset paper states that a token becomes eligible for Raydium once its Market Capitalization reaches \$69,000, and that only tokens that reached this threshold and moved to Raydium were retained in the resulting corpus (Long et al., 2024). Another defines “viability” directly by whether the memecoin was successfully listed on Raydium, again treating the Pump.fun-to-Raydium transition as the salient event (Long et al., 2024). These differing formulations are paper-specific descriptions of Pump.fun’s launch-to-listing pathway across different empirical settings.

A central implication of the formal market-design analysis is that price continuity at graduation does not imply liquidity continuity. The marginal price is constructed to be continuous when the token migrates, but the post-graduation pool is shallower because the synthetic liquidity disappears. The paper shows that, for any fixed inventory size, selling just before graduation yields more SOL than selling immediately after, which creates a strong incentive for early large holders to dump before graduation (Marino et al., 16 Feb 2026).

3. Pump.fun as a research corpus: datasets and observables

Pump.fun’s standardized page structure and program-level traceability have made it a primary source of multimodal memecoin datasets. The “Coin-Meme” dataset is built entirely from Pump.fun tokens that successfully moved to Raydium between January 2024 and November 2024. Using Pump.fun’s Solana program address, 6EF8rrecthR5Dkzon8Nwu78hRvfCKubJ14M5uBEwF6P, the authors extracted token addresses, creation times, names, and creators from Dune.com, then scraped token descriptions, images, user comments, and financial data from URLs of the form https://pump.fun/<token_address>. After filtering for Pump.fun memecoins that subsequently moved to Raydium, the final dataset contains 3,751 memecoins. Text was lowercased, stripped of non-alphabetic characters, stopwords were removed, and text was tokenized and lemmatized; entries with missing descriptions or comments were excluded. Images were resized to 224×224, non-RGB images were converted to RGB, and corrupted or missing images were replaced with zero vectors. The dataset captures four modalities per token: textual descriptions, visual content, community interactions, and financial data, including Market Capitalization and Market Entry Time, where Market Entry Time is the time difference between Pump.fun creation and movement to Raydium (Long et al., 2024).

The “CoinVibe” dataset is also built explicitly from Pump.fun memecoins, but it frames the task as viability classification rather than thematic clustering. It contains 6,231 memecoins from January 2024 to November 2024 and includes textual descriptions, images standardized to 224×224 RGB for CLIP’s ViT-L/14 encoder, and community data consisting of user comments, timestamps, and number of likes. Labels are binary: tokens listed on Raydium are “Viable,” and tokens not listed are “Non-Viable.” The paper reports a class distribution of 44.27% viable and 55.73% non-viable, with predefined 80%/10%/10% train/validation/test splits. As in Coin-Meme, tokens missing key modalities such as descriptions or comments were excluded during data cleaning (Long et al., 2024).

Taken together, these datasets represent Pump.fun as a standardized empirical environment in which the memecoin is not merely a token contract but a multimodal object composed of narrative text, logo imagery, user discourse, and protocol-defined financial milestones. A plausible implication is that Pump.fun’s research value derives as much from its interface regularity as from its market scale.

4. Multimodal analysis and the cultural structure of Pump.fun memecoins

The Coin-Meme study proposes a multimodal framework in which each Pump.fun memecoin is encoded through textual, visual, and financial features and then clustered after dimensionality reduction. Token descriptions are modeled with Latent Dirichlet Allocation, producing topic-proportion vectors

$\mathbf{t}_i = (p_{i1}, p_{i2}, \dots, p_{iK}),$

while logo images are passed through ResNet50 to obtain high-dimensional embeddings

$\mathbf{v}_i \in \mathbb{R}^d.$

Financial features consist of scaled Market Entry Time and Market Capitalization,

$\mathbf{f}_i = (\tilde{E}_i,\tilde{\text{MC}}_i),$

and the unified representation is the concatenation

$x\,y = k, \qquad P = \frac{x}{y}.$0

PCA is applied before K-Means clustering, and the number of clusters is selected by silhouette score. The final solution contains three clusters: Cluster 0 with 1,127 coins, Cluster 1 with 1,535 coins, and Cluster 2 with 1,089 coins (Long et al., 2024).

The three clusters correspond to distinct cultural themes. Cluster 0 is centered on humor, parody, and niche internet culture, with word-cloud terms such as “meme” and “pepe,” and logos featuring exaggerated imagery, caricatures, and fantastical creatures. Cluster 1 is organized around animals, cuteness, and whimsical themes, with words such as “cat,” “dog,” “cutest,” “friend,” and “fun,” and logos depicting cats, dogs, tigers, red pandas, and meme animals such as Grumpy Cat. Cluster 2 centers on political satire and iconic figures, especially Donald Trump, with word-cloud terms such as “trump,” “crypto,” and “world,” and logos dominated by caricatural portrayals of authority figures. Across all clusters, frequent shared terms include “world,” “community,” and “meme,” indicating recurrent strategies of global scope, collective identity, and explicit meme signaling (Long et al., 2024).

The same study characterizes community interaction through comments. Sentiment analysis uses a Transformers sentiment-analysis pipeline with BERT, producing positive-to-negative comment ratios, per-comment sentiment scores, and sentiment variability

$x\,y = k, \qquad P = \frac{x}{y}.$1

Comments per user are computed as total comments divided by unique users. At the cluster level, the positive-to-negative comment ratio is 1.23 for Cluster 0, 1.36 for Cluster 1, and 1.64 for Cluster 2; comments per user are 2.97, 3.74, and 2.78 respectively; sentiment variability is 0.3761, 0.3754, and 0.3394 respectively. The authors note that the relatively low sentiment variability in Cluster 2 may reflect more controlled or targeted bot activity. This does not establish manipulation conclusively, but it identifies a measurable asymmetry between political-satire coins and the other thematic groups (Long et al., 2024).

5. Predictive modeling of viability and graduation

One research line treats Pump.fun memecoins as a multimodal classification problem. CoinCLIP uses frozen CLIP encoders—ViT-L/14 for images and a Transformer-based text encoder—for logo and description embeddings, followed by modality-specific projection layers, separate feature adapters with residual connections, and a community-data pathway that encodes comments with the CLIP text encoder and transforms normalized timestamps and like counts into embeddings. Per-comment features are weighted by likes and aggregated into a community representation $x\,y = k, \qquad P = \frac{x}{y}.$2. Image and text features are fused through a Hadamard product,

$x\,y = k, \qquad P = \frac{x}{y}.$3

then concatenated with $x\,y = k, \qquad P = \frac{x}{y}.$4 and classified by an MLP trained with cross-entropy for binary viability prediction. On the CoinVibe dataset, the reported performance is 84.72$x\,y = k, \qquad P = \frac{x}{y}.$5 accuracy, 92.07$x\,y = k, \qquad P = \frac{x}{y}.$6 AUROC, and 83.74$x\,y = k, \qquad P = \frac{x}{y}.$7 macro-F1, outperforming BERT, CLIP Text-Only, ViT-L/14, CLIP Image-Only, CLIP, and CLIP-Adapter. The paper further reports that image-only models outperform text-only models, and that adding community data improves all metrics in the ablation analysis (Long et al., 2024).

A complementary descriptive analysis links thematic clusters to financial behavior. In Coin-Meme, Cluster 0 has the slowest Mean Market Entry Time at 25,897 seconds and the highest 95% quantile of Market Capitalization at \$x\,y = k, \qquad P = \frac{x}{y}.$$824,970.5; Cluster 2 has the fastest Mean Market Entry Time at 4,728 seconds and a 95$$29,508.5. The authors summarize these profiles as slower-liquidity but higher-upside “niche value” for humor/parody coins, faster adoption but lower top-end capitalization for animal-themed coins, and rapid early traction with moderate top-end capitalization for political-satire coins (Long et al., 2024).

A different predictive framework models Pump.fun graduation directly from on-chain state variables rather than from page-level multimodal content. The central quantity is the empirical conditional graduation probability as a function of virtual SOL in the bonding curve, $x\,y = k, \qquad P = \frac{x}{y}.$9. The baseline curve

$y_0^{\rm mint} = 10^9 \text{ tokens},$0

is strictly increasing and approaches 1 near the graduation threshold. The study then conditions this curve on four classes of variables: share of non-bot trades, number of trades required to reach a given $y_0^{\rm mint} = 10^9 \text{ tokens},$1, early participation by historically profitable wallets, and launch by top creators. The strongest predictor is fast liquidity accumulation in a small number of trades: for a fixed $y_0^{\rm mint} = 10^9 \text{ tokens},$2, tokens that reach that level in few trades have the highest graduation probability. A higher share of non-bot participation also raises graduation odds, while top-trader presence provides a positive but modest uplift and top-creator effects can exceed the breakeven boundary at sufficiently high $y_0^{\rm mint} = 10^9 \text{ tokens},$3, albeit with noisy estimates. The paper introduces a simple breakeven benchmark,

$y_0^{\rm mint} = 10^9 \text{ tokens},$4

to evaluate whether a naive buy-and-hold-until-graduation rule would have positive expected return; baseline graduation probabilities remain below this curve throughout, indicating that $y_0^{\rm mint} = 10^9 \text{ tokens},$5 alone is not sufficient for profitable naive timing (Marino et al., 16 Feb 2026).

6. Risks, controversies, and broader significance

The literature consistently portrays Pump.fun as a high-churn, high-risk environment. In the Q4 2024 Solana study, the daily graduation rate peaks at less than 2%, meaning that over 98% of created tokens fail to reach the Raydium listing condition. The same study interprets the combination of 40% to 67.4% of DEX trades involving Pump.fun-created tokens, volume share that rarely exceeds 50%, and rapid growth in daily active users as evidence of a retail-dominated market characterized by frequent small trades and short-term speculation rather than long-term value creation. It also notes controversy around Pump.fun’s livestream feature, which was indefinitely suspended after criticism that some token creators used extreme actions to promote their tokens (Mancino, 4 Dec 2025).

The on-chain market-design study sharpens this risk characterization. Over the period 1 September to 1 October 2025, 655,770 tokens were created by 243,123 distinct creator addresses, but only 4,338 graduated, a graduation rate of about 0.63%. Among 184,282 tokens with at least 30 swaps, 92.22% exhibited at least one statistically large negative price shock under a Shewhart-style dump rule, and only 2.55% of those graduated. The study attributes much of this failure to pre-graduation dump incentives created by the liquidity discontinuity at migration: early stakeholders can often realize more SOL by exiting just before graduation than by waiting until after it (Marino et al., 16 Feb 2026).

At the same time, Pump.fun is not described merely as a venue for arbitrary fraud. The Q4 2024 study emphasizes fair-launch design, the absence of presales and team allocations, and permanently burned Raydium liquidity for graduated tokens as mechanisms intended to reduce classic rug-pull patterns (Mancino, 4 Dec 2025). The 2026 market-design paper adds that creators cannot rug-pull the virtual pool because liquidity cannot be added or removed manually in the virtual AMM; only standard swaps are allowed (Marino et al., 16 Feb 2026). A common misconception is therefore that Pump.fun’s standardization eliminates strategic abuse. The literature rejects that view: standardization appears to reduce some contract-level exploit classes while leaving pump-and-dump dynamics, bot activity, and adverse selection fully salient.

Methodological limits are also explicit. Coin-Meme contains only tokens that successfully moved to Raydium, so it is affected by selection bias toward relatively successful memecoins. CoinVibe defines viability narrowly as Raydium listing, which excludes long-term post-listing outcomes and may underrepresent other notions of success. Both datasets cover January to November 2024 and may therefore be sensitive to temporal drift as meme formats, platform policies, and market conditions change. CoinCLIP is Pump.fun-centric and may not generalize directly to other chains or launchpads, while its reliance on logo, comment, and like patterns may bias it toward historically successful meme aesthetics and pre-existing communities (Long et al., 2024).

Taken together, these studies depict Pump.fun as a dual-purpose institution within Web3. It democratizes token creation, aggregates retail attention at unprecedented scale, and yields unusually rich multimodal and on-chain data for studying the conversion of cultural symbols into financial instruments. It also creates a market structure in which success is rare, attention is endogenous to interface design and community signaling, and incentive-compatible speculation can dominate project durability. This suggests that Pump.fun is best understood not simply as a memecoin website, but as a standardized socio-financial mechanism linking meme production, bonding-curve price discovery, and selective migration into the broader Solana DeFi ecosystem.