HoloByte: Tokenizer-Free Sequence Modeling

• HoloByte is a tokenizer-free sequence modeling framework that converts fixed-width byte chunks into continuous hyperspherical vectors via invertible orthogonal rotations.
• It employs macroscopic Transformers with local causal micro-decoders to achieve efficient, exact byte-level recovery while reducing computational complexity.
• The framework’s design inspires diverse holographic applications, from opto-magnetic storage and magnonic memory to simulated holographic telepresence.

Searching arXiv for the cited HoloByte-related papers and closely related context papers to ground the article. HoloByte is the name of a tokenizer-free sequence-modeling framework introduced in 2026, in which fixed-width chunks of raw bytes are compressed into continuous, strictly bounded hyperspherical vectors by an invertible orthogonal rotation operator, processed by a macroscopic Transformer, and locally decoded back into exact byte-level distributions by a causal micro-decoder (Khasia, 10 Mar 2026). In adjacent holography literatures, closely related “HoloByte-style” usages designate information-bearing holographic cells or systems rather than a language-model architecture: dynamic complex opto-magnetic holography uses a detour-phase Lohmann cell as a “holographic byte,” magnonic holographic memory describes holography-based 8-bit storage via spin-wave interference, and a simulated holographic telepresence system uses multi-view point-cloud fusion and a quadrangular acrylic pyramid for omnidirectional visualization (Makowski et al., 2022).

1. Scope and terminology

Within the literature represented here, “HoloByte” is not a single-domain term. Its most precise and explicit use is the sequence-modeling framework “HoloByte: Continuous Hyperspherical Distillation for Tokenizer-Free Modeling” (Khasia, 10 Mar 2026). Two hardware-oriented usages are structurally analogous rather than identical: one defines a detour-phase holographic cell that stores amplitude and phase, and another explains how a magnonic holographic memory can realize a holography-based 8-bit data unit (Makowski et al., 2022). A further “HoloByte-style” usage appears in simulated holographic telepresence, where the emphasis is not storage or tokenization but omnidirectional 3D rendering from RGB-D fusion (Córdova-Esparza et al., 2018).

Usage	Substrate	Core mechanism
HoloByte	Sequence modeling	Continuous Hyperspherical Distillation
“Holographic byte”	Opto-magnetic holography	Detour-phase Lohmann cell with amplitude and phase
“HoloByte” in memory	Magnonic holographic memory	Spin-wave interference modulated by nanomagnets
HoloByte-style telepresence	Simulated holographic display	Multi-camera 3D reconstruction and four-view pyramid rendering

This terminological spread matters because the shared name does not imply a shared implementation. In the machine-learning setting, HoloByte is a tokenizer-free autoregressive architecture over bytes. In the holography settings, the common denominator is compact information encoding in a physically reconstructed field or display. This suggests a family resemblance centered on compressed, directly decodable representations, but the underlying mathematics and hardware differ substantially.

2. Continuous Hyperspherical Distillation architecture

The 2026 HoloByte framework begins from the standard byte vocabulary $\mathcal{V} = \{0,1,\dots,255\}$ and partitions a byte sequence $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$ into contiguous chunks of width $W$, with $T=\lfloor N/W\rfloor$ chunks $c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$ (Khasia, 10 Mar 2026). Each byte $v$ is mapped to a learnable embedding $m_v \in \mathbb{R}^D$, then normalized onto the unit hypersphere,

$$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$

The use of $\mathbb{S}^{D-1}$ is central: the framework is explicitly continuous, strictly bounded, and vocabulary-invariant at the macroscopic level.

Intra-chunk position is encoded by a position-dependent orthogonal operator $\mathcal{R}_i \in \mathrm{SO}(D)$. For even $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$0, if $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$1 with $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$2, frequencies $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$3, and angles $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$4, then

$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$5

This rotation is an exact isometry: $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$6, $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$7, and $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$8. Distances and inner products are preserved under rotation.

Chunk compression is performed by “spatial superposition,” which binds $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$9 unit-norm byte signals into a single $W$0-dimensional vector:

$W$1

The resulting macroscopic sequence $W$2 is then processed by a causal self-attention model $W$3,

$W$4

so the Transformer operates on length $W$5 rather than on the original byte length $W$6.

Exact byte-level recovery is delegated to a localized causal micro-decoder. First, HoloByte unbinds each position by inverse rotation,

$W$7

Strict causality is enforced by a right-shifted prefix,

$W$8

A single-layer causally masked self-attention micro-decoder $W$9 operates on the local sequence $T=\lfloor N/W\rfloor$0 and outputs $T=\lfloor N/W\rfloor$1. Byte-level logits are then computed by normalized cosine similarity against the hyperspherical manifold with temperature $T=\lfloor N/W\rfloor$2:

$T=\lfloor N/W\rfloor$3

The paper initializes $T=\lfloor N/W\rfloor$4, so $T=\lfloor N/W\rfloor$5, in FP32.

3. Objectives, complexity, and theoretical guarantees

The motivating claim of HoloByte is that subword tokenization avoids the $T=\lfloor N/W\rfloor$6 cost of native byte-level attention at the price of artificial morphological boundaries, vocabulary dependence, and a fractured optimization landscape (Khasia, 10 Mar 2026). HoloByte keeps byte-level vocabulary invariance while reducing the effective sequence length seen by the macroscopic Transformer from $T=\lfloor N/W\rfloor$7 to $T=\lfloor N/W\rfloor$8. The resulting total complexity is stated as

$T=\lfloor N/W\rfloor$9

with peak attention memory

$c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$0

The first term reflects macro attention on the shortened sequence; the second reflects feed-forward and micro-decoder costs. Because micro passes are chunk-local and can be batched over $c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$1, the design trades global quadratic cost for a shorter-sequence macro model plus fixed-width local decoding.

Training uses a dual-objective formulation. Cross-entropy is applied directly to byte distributions,

$c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$2

while the Holographic Latent Mean Squared Error uses the deterministic next-chunk target $c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$3:

$c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$4

The total loss is

$c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$5

The latent term is not auxiliary in a vague sense; it is presented as a mathematically precise restorative force that pulls $c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$6 into a bounded ball of radius $c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$7. Its gradient is

$c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$8

and the paper gives the bound

$c_t = (b_{t,0},\dots,b_{t,W-1}) \in \mathcal{V}^W$9

with a Lipschitz constant of order $v$0 for the latent dynamics. The stated significance is asymptotic stability and gradient control independent of $v$1.

The main recoverability guarantee concerns interference after unbinding. Under isotropic distribution of manifold vectors on $v$2, expected interference remains $v$3, and decoding by margin $v$4 with failure probability $v$5 requires

$v$6

The abstract gives the corresponding simplified statement $v$7 for error-free discrete recovery (Khasia, 10 Mar 2026). Since $v$8 for bytes, the theory links chunk width $v$9, embedding dimension $m_v \in \mathbb{R}^D$0, and recoverability in a direct way: larger $m_v \in \mathbb{R}^D$1 improves macro compression but raises interference and therefore raises the required $m_v \in \mathbb{R}^D$2.

4. Empirical behavior and implementation profile

The reported empirical study uses a FineWeb-Edu subset of approximately $m_v \in \mathbb{R}^D$3 characters, deterministic hardware and seed, and a standard autoregressive Transformer backbone under parameter parity of approximately $m_v \in \mathbb{R}^D$4M parameters (Khasia, 10 Mar 2026). The baseline BPE system uses a GPT-2 vocabulary of $m_v \in \mathbb{R}^D$5, $m_v \in \mathbb{R}^D$6, and $m_v \in \mathbb{R}^D$7 macro layers, with the large embedding table occupying approximately $m_v \in \mathbb{R}^D$8M parameters. The HoloByte configuration uses $m_v \in \mathbb{R}^D$9, $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$0, $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$1, $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$2 macro layers, and $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$3 micro layer; its byte-embedding table is approximately $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$4M parameters, leaving more capacity for depth.

The principal metric is absolute nats per byte. For the BPE baseline, nats per token are converted using average bytes-per-token $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$5,

$$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$6

At $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$7k steps, the baseline attains $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$8, yielding approximately $$\tilde{\mathbf{m}}_v = \frac{\mathbf{m}_v}{\|\mathbf{m}_v\|_2} \in \mathbb{S}^{D-1}.$$9 nats/byte. HoloByte’s total validation loss converges to $\mathbb{S}^{D-1}$0, implying $\mathbb{S}^{D-1}$1 nats/byte. The training curves are described as monotonic, and the framework is reported as systematically outperforming a comparable discrete BPE baseline under strictly matched parameter constraints (Khasia, 10 Mar 2026).

The implementation profile is correspondingly specific. Typical hyperparameters are $\mathbb{S}^{D-1}$2, $\mathbb{S}^{D-1}$3, $\mathbb{S}^{D-1}$4, learning rate $\mathbb{S}^{D-1}$5, AdamW, weight decay $\mathbb{S}^{D-1}$6, gradient clip $\mathbb{S}^{D-1}$7, and AMP for matmuls with FP32 retained for normalization and cosine computations. The rotation operator is deterministic and parameter-free. Training uses teacher forcing for the micro-decoder; inference autoregresses within each chunk with right padding up to $\mathbb{S}^{D-1}$8. The code repository is provided at https://github.com/VladimerKhasia/HoloByte (Khasia, 10 Mar 2026).

The reported ablations are qualitative but structurally important. Increasing $\mathbb{S}^{D-1}$9 improves macro attention savings proportional to $\mathcal{R}_i \in \mathrm{SO}(D)$0 but increases micro cost proportional to $\mathcal{R}_i \in \mathrm{SO}(D)$1 and raises interference. Increasing $\mathcal{R}_i \in \mathrm{SO}(D)$2 reduces interference variance proportional to $\mathcal{R}_i \in \mathrm{SO}(D)$3 and better satisfies the recoverability bound, but increases $\mathcal{R}_i \in \mathrm{SO}(D)$4 cost. Orthogonal rotation preserves isometry and exact invertibility, hyperspherical normalization stabilizes cosine logits, and removing the latent term increases gradient variance from cross-entropy alone. These points are not merely implementation choices; they are part of the framework’s claim to be both continuous and exactly decodable.

5. Holographic storage, memory, and display interpretations

In dynamic complex opto-magnetic holography, a “HoloByte” is not a byte sequence model but a detour-phase Lohmann cell that stores two degrees of freedom: amplitude, encoded by the circular opening diameter $\mathcal{R}_i \in \mathrm{SO}(D)$5, and phase, encoded by the spatial detour $\mathcal{R}_i \in \mathrm{SO}(D)$6 (Makowski et al., 2022). The physical medium is a transparent ferrimagnetic stack, AlTi(10 nm)/Si$\mathcal{R}_i \in \mathrm{SO}(D)$7N$\mathcal{R}_i \in \mathrm{SO}(D)$8(5 nm)/Gd$\mathcal{R}_i \in \mathrm{SO}(D)$9Fe$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$00Co$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$01(20 nm)/Si$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$02N$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$03(60 nm) on glass, $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$04 mm, with perpendicular magnetization and coercivity approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$05 Oe. All-optical switching uses a single $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$06 fs pulse at $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$07 nm, with a toggle-ready state in approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$08 ps. Although the write beam has a Gaussian spot of approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$09 at $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$10, threshold switching yields domains as small as $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$11. The relevant encoding relations are

$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$12

with $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$13. The system writes each computed cell immediately, without materializing a full hologram in RAM. Demonstrated patterns include $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$14 points and complex-encoded patterns with $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$15 Lohmann cells. With a $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$16 GHz laser and single-beam serial write, the paper estimates approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$17 million points per frame at $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$18 Hz; with four MEMS/FPGA/beam units, approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$19 million points at $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$20 Hz.

In magnonic holographic memory, the relevant entity is an 8-bit data unit realized through spin-wave interference in a magnetic matrix modulated by nanomagnets (Gertz et al., 2014). The demonstrated device uses a YIG film on GGG substrate with thickness $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$21, $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$22 G, and FMR linewidth $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$23 Oe. The structure length is $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$24 mm with arm width $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$25, six micro-antennas for I/O, and two top micro-magnets as memory elements. Operation occurs in the $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$26–$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$27 GHz range, with data shown near $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$28–$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$29 GHz and bias fields up to $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$30 Oe. The spin-wave field at a detector is written as

$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$31

A concrete 8-bit scheme sets $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$32 inputs and encodes each bit by binary phase-shift keying,

$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$33

The paper emphasizes room-temperature coherence, robust interference over $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$34 mm paths, and projected areal density up to approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$35 Tb/cm$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$36, although it also notes that practical density will be limited by I/O element size, damping, and fabrication tolerances.

A distinct display-oriented usage appears in “Telepresence System based on Simulated Holographic Display,” which describes a HoloByte-style telepresence pipeline rather than a data unit (Córdova-Esparza et al., 2018). The system uses four Kinect V2 sensors placed at approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$37 meters in height with viewpoint changes of about $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$38, reconstructs a colored 3D point cloud by multi-view fusion, and renders four virtual views onto a quadrangular acrylic pyramid using a commodity projector. Calibration uses a 1D wand with three collinear points, lens distortion is modeled with the Brown model, and nonlinear refinement uses Levenberg–Marquardt. Reported mean reprojection error is $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$39 pixels for depth cameras and $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$40 pixels for color cameras. Foreground segmentation uses an encoder–decoder CNN trained on CDNet2014 with overall Recall $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$41, Specificity $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$42, Precision $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$43, and F1 $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$44. Box reconstruction errors are approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$45 mm in width and $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$46 mm in height, and the end-to-end runtime is approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$47 FPS. This usage is holographic in the display sense: the display produces an omnidirectional visualization effect reminiscent of Pepper’s Ghost projection without head-mounted displays.

6. Limitations and open directions

For the tokenizer-free HoloByte model, the limitations are internal to the continuous compression mechanism. The paper states that maintaining exact unit normalization and FP32 numerical stability is important; drift or loss of orthogonality would degrade recovery (Khasia, 10 Mar 2026). Interference grows with $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$48, so aggressive compression requires larger $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$49 in line with the dimension bound. The micro-decoder adds per-chunk overhead even though it is parallelizable, and a single micro layer may be insufficient for extremely complex intra-chunk dependencies in some domains. The proposed practical guidance is to use $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$50 in the range $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$51–$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$52, with $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$53 as a default, and to choose $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$54 according to

$b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$55

for $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$56 (Khasia, 10 Mar 2026).

For dynamic complex opto-magnetic holography, the bottlenecks are physical rather than statistical. Write/erase fatigue and domain wall motion can transiently degrade contrast; the paper reports that domain wall motion becomes negligible after approximately four write-reset cycles (Makowski et al., 2022). Current toggled domains are approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$57 with low-NA optics, finer domains require higher NA or engineered media, and parallel data transfer and writing optics remain a systems challenge for the target of $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$58 cells at $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$59 Hz. The paper identifies materials engineering, many-beam parallelization, higher FPGA clock rates, adaptive carriers, and multi-color operation as future work.

For magnonic holographic memory, the principal constraints are damping, phase stability, parasitic coupling, and I/O scaling (Gertz et al., 2014). Ferrite flicker noise is approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$60 dBm, but parasitic inter-port coupling and reflections of approximately $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$61 to $b = (b_0,b_1,\dots,b_{N-1}) \in \mathcal{V}^N$62 dBm dominate residual signals at nominal destructive interference. Micro-antennas scale poorly because induced voltage scales with loop area, and the paper therefore points to spin torque nano-oscillators and magnetoelectric elements as alternatives for scaling and lower cross-talk. Geometry choice between MSSW and BVMSW affects phase sensitivity and defect scattering, and CMOS-compatible RF control remains an integration challenge.

For simulated holographic telepresence, limitations arise from sensing, calibration, and display optics (Córdova-Esparza et al., 2018). The representation is a visual-only colored point cloud rather than a temporally consistent mesh with high-fidelity texturing. The segmentation model is explicitly reported as not state of the art, especially in turbulence and night videos, and the static camera assumption excludes PTZ or jitter scenarios. Sensor noise, occlusions, projector reflections, ambient lighting, and extreme viewing angles all degrade the perceived volumetric effect. Future work in the paper includes temporally consistent 3D meshes, higher-fidelity audio transmission, improved sensors, advanced fusion methods such as TSDF or Poisson reconstruction, scalable multi-site streaming, and better display optics.

Taken together, these literatures show that HoloByte names either a specific tokenizer-free continuous representation for byte sequences or a class of compact holographically encoded physical units and systems. The shared theme is exact or calibrated recovery from compressed superposition: hyperspherical superposition and micro-decoding in sequence modeling, amplitude-and-phase detour encoding in opto-magnetic media, interference reconstruction in magnonic memory, and multi-view projection in simulated holographic telepresence.