Little red dots.
The search is ongoing.

JWST recorded this light when the universe was six hundred sixty-four million years old. The same patch of sky, band by band. Barely there in the blue. In the red, a point.

Ten filters, laid out by wavelength. Each point is how much light arrives in one filter. Together they trace the object's color across the infrared. Faint blue light, then a sharp rise into the red. Between these two bands, the flux jumps nearly tenfold.

A spectrum now, instead of ten samples of one. Every wavelength, in order. Hydrogen absorbs hard at one particular wavelength, but only when its electrons are already excited. That takes dense, warm gas. Stars can make a break, but the stellar templates compared here do not make one this large.

Here is what the model puts inside. At the center, a black hole of about two million solar masses, swallowing gas. Around it, the last orbit gas can hold before it falls. Then a dense envelope of gas surrounds the center. Light escapes only at the photosphere, where the gas finally thins. Everything beneath it is inference.

Consider a photon trying to escape through a layer of ionized gas. In this example, free electrons scatter it again and again. Forty-two optical depths imply about seventeen hundred scatterings. Across this model's skin: six hours in a straight line, about ten days by diffusion. Different routes spread out the arrival times. The delay depends on the layer's thickness and the assumed geometry.

So how heavy is it? That turns on one number: the width of a hydrogen line. There it is, on the same spectrum. About three thousand kilometers a second. Line widths can trace motion. Read this one as orbits, and the estimate is two hundred million suns. Resonant scattering by hydrogen offers another way to broaden that line. Then it is around a million instead, running five times past its Eddington limit.

Start with a seed, and grow it to the mass inferred at the time we observe it. The heavy example takes one and a half billion years. The universe here is only six hundred sixty-four million. The lighter example takes ninety-two million years. Both assume a seed mass, steady feeding, and a fixed efficiency. Change those assumptions and the times change. The clock constrains a growth history; it does not choose the model.

If that is what this one is, it may be what a whole population is. Little red dots present several puzzles. This model offers a possible connection between their spectrum, size and weak X-rays. Extending the fitted blue spectrum into the red leaves an excess. Calling that excess an engine is the model. That is one possible reading.

English captions (VTT) ↗

At redshift 7.7569, we see this object when the universe was 664 million years old (Planck18 cosmology). An illustrative blackbody at an assumed 7,500 K has a radius of about 792 AU, or a diameter of about 1,600 AU, for the model luminosity of 10^^44.5 erg/s. This is a model scale, far below the image resolution.

One line width, different mass estimates

The published Nature paper explores how the line width changes the inferred mass. Standard orbital scaling gives about 10^^8.3 solar masses; omitting the dust correction gives about 10^^7.7. The resonant Balmer-scattering interpretation gives about 10^⁶. These are alternative, model-dependent readings, not three independent mass measurements (published Extended Data Table 4; Table ED 5 in the earlier preprint).

Which luminosity? The Moon and size comparisons use the integrated atmosphere-model luminosity, 10^^44.5 erg/s. The paper’s Extended Data Table 4 gives different, Hβ-based estimates: log luminosity 45.7 with dust correction (quoted Eddington ratio 0.25), 45.0 without it (0.16), and approximately 45 for the scattering example (approximately 5). These are alternative assumptions with rounded estimates.

For a direct check, use L_Edd = 1.26 × 10^³⁸ × (M / M☉) erg/s. Holding the atmosphere luminosity fixed gives L / L_Edd ≈ 0.013 at 10^^8.3 solar masses, or 2.5 at 10^⁶. The growth examples below instead use the paper’s quoted ratios; they are illustrative histories, not ratios calculated from the Moon comparison. The rounded table entries do not give exact ratios when divided.

The growth comparison must build each inferred mass before the observation. With radiative efficiency 0.1 and constant feeding, a 10^⁵-solar-mass seed takes about 1.52 billion years to reach 10^^8.3 at 0.25 times Eddington. A 10^²-solar-mass seed takes about 92 million years to reach 10^⁶ at five times Eddington. These examples depend on the seed, efficiency and feeding history; seed formation also uses part of the universe’s first 664 million years.

The blue curve in the last panel is fitted to this object’s ultraviolet photometry: β = −1.25 ± 0.28. Dotted segments extrapolate that fit. Interpreting the blue light as a host galaxy and the red excess as a central engine is a model, rather than a measured decomposition.

Selection matters. Rinaldi et al. identify 412 sources across redshift 2–11; extreme color cuts recover about a quarter or less of that candidate sample. Weibel et al. identify 241 black-hole-star candidates spanning redshift 1.7–9.3, an interval of about 3.3 billion years in cosmic time.

The wider class need not share one explanation. Kiyota et al. find evidence for dust in at least about half their studied sample. Yanagisawa et al. report blueshifts in 38 of 46 Balmer absorbers. Higher-resolution line profiles can test the atmospheric interpretation.

If central-engine light is fitted as starlight, inferred stellar masses can be biased upward. That possibility alone does not resolve the abundance of massive galaxies at early times.

• JWST positions: the twelve source catalogs below. Their original selections are retained.
• Euclid: Bisigello et al., Q1 LRD sample, A&A 711, A24, via VizieR J/A+A/711/A24; Tumborang et al., COSMOS sample, the 16 LRD/LBD candidates in Table 3. DESI: Lin et al., DR1 catalog.
• MoM-BH*-1: Naidu et al., Nature (2026) · preprint. Spectra: Dawn JWST Archive; aperture photometry measured from released DJA/PRIMER frames.

The map uses the releases listed here. Entries can repeat the same object; paper revisions and later catalog releases may have different totals. Year labels follow the catalog labels; these may use the preprint or journal-publication year.

Catalog references and input checksums (JSON) · Original nine references checked 13 September 2026; three catalog additions checked 17 September 2026.