Roblox Binary Model Format, Version 0

Roblox DOM and (de)serialization implementation in Rust


Roblox Binary Model Format, Version 0

This is unofficial documentation for Roblox’s binary model format. The binary model format is used for places (.rbxl files), models (.rbxm files), and many objects uploaded to Roblox’s asset storage.

The binary model format intended to supersede Roblox’s older XML model format.

This document is based on:

Contents

Document Conventions

This document assumes a basic understanding of Rust’s conventions for numeric types. For example:

  • u16 is an unsigned 16-bit integer
  • i32 is a signed 32-bit integer

Integers are assumed to be little endian and 2’s complement unless otherwise specified. The presence of big endian integers and integers with interesting transformations are explicitly noted.

The data contained in a chunk may be compressed. The term “chunk data” refers to the decompressed contents.

File Structure

Binary model files consist of a short header, followed by a series of chunks. Each chunk has the same framing, enabling consumers to partialy decode a file.

  1. File Header
  2. Chunks
    1. Zero or one META chunks
    2. Zero or one SSTR chunks
    3. Zero or more INST chunk
    4. Zero or more PROP chunks
    5. One PRNT chunk
    6. One END chunk

File Header

Every file starts with a 32 byte header.

Field Name Format Value
Magic Number 8 bytes Always <roblox!
Signature 6 bytes Always 89 ff 0a 1a 0a
Version u16 Always 0
Class Count i32 Number of distinct classes in the file (i.e. the number of INST chunks)
Instance Count i32 Number of instances in the file
Reserved 8 bytes Always 0

Chunks

Every chunk starts with a 16 byte header followed by the chunk’s data.

Field Name Format Value
Chunk Name 4 bytes The chunk’s name, like META or INST
Compressed Length u32 Length of the chunk in bytes, if it is compressed
Uncompressed Length u32 Length of the chunk’s data after decompression
Reserved 4 bytes Always 0

If Chunk Name is less than four bytes, the remainder is filled with zeros.

If Compressed Length is zero, Chunk Data contains Uncompressed Length bytes of data for the chunk.

If Compressed Length is nonzero, Chunk Data contains an LZ4 compressed block. It is Compressed Length bytes long and will expand to Uncompressed Length bytes when decompressed.

When the chunk data is compressed, it is done so using the LZ4 compression algorithm.

META Chunk

The META chunk has this layout:

Field Name Format Value
Number of Metadata Entries u32 The number of metadata entries in the chunk
Metadata Entries Array(Entries) The actual metadata entries

Each metadata entry has the following format:

Field Name Format Value
Metadata Key String The metadata key, which should be unique
Metadata Value String The value for this metadata key

The Metadata chunk (META) is a map of strings to strings. It represents metadata about the model, such as whether it was authored with ExplicitAutoJoints enabled.

There should be zero or one META chunks.

Observed metadata entries and their values:

  • ExplicitAutoJoints: true or false

SSTR Chunk

The SSTR chunk has this layout:

Field Name Format Value
Version u32 The version of the SSTR chunk (always 0)
Shared String Count u32 The number of SharedStrings in the chunk
Strings Array(Shared Strings) The actual shared string entries

A shared string entry looks like this:

Field Name Format Value
MD5 Hash 16 bytes An MD5 hash of the Shared String
Shared String String The string that’s used by a later PROP chunk

The Shared String chunk (SSTR) is an array of strings. It’s used to reduce the overall size of a file by allowing large strings to be reused in PROP chunks. The MD5 Hash isn’t used by Roblox Studio when loading the file.

There should be zero or one SSTR chunks.

INST Chunk

The INST chunk has this layout:

Field Name Format Value
Class ID u32 An arbitrarily-chosen ID referring to a Roblox class
Class Name String The class name, like Folder or Part
Object Format u8 1 if the class is a service, otherwise 0
Instance Count u32 The number of instances belonging to the class
Referents Array(Referent) The referents of instances belonging to the class
Service Markers Array(u8) 1 for each instance if the class is a service, otherwise not present

There should be one INST chunk for each type of instance defined.

There are two forms of the INST chunk determined by the Object Format field:

  • 0: regular
  • 1: service

If the Object Format is regular, the service markers section will not be present.

If the Object Format is service, the service markers section contains 1 repeated for the number of instances of that type in the file. If this field is not set, Roblox may create duplicate copies of services, like in rojo-rbx/rbx-dom#11.

Class ID must be unique and ideally sorted monotonically among all INST chunks. It’s used later in the file to refer to this type.

Class Name should match the ClassName specified on an instance in Roblox.

The length of Referents must equal Instance Count.

PROP Chunk

The PROP chunk has this layout:

Field Name Format Value
Class ID u32 The class ID assigned in the INST chunk
Property Name String The name of the property, like CFrame
Type ID u8 The Data Type of the property
Values Array(Value) A list of values whose type is determined by Type ID

The property chunk (PROP) defines a single property for a single instance type.

There should be one PROP chunk per property per instance type.

Because of the shape of this chunk, every instance of a given class must have the same properties specified with the same times. Put another way, if an instance in the file defines a property, all other instances of the same class must also specify that property!

Class ID defines the class that this property applies to as defined in a preceding INST chunk.

Property Name defines the serializable name of the property. Note that this is not necessarily the same as the name reflected to Lua, which is sometimes referred to as the canonical name.

Type ID corresponds to a Data Type’s Type ID.

Values contains an array of values whose type is determined by Type ID and whose length is equal to the number of instances belonging to Class ID.

PRNT Chunk

The PRNT chunk has this layout:

Field Name Format Value
Version u8 Always 0
Instance Count u32 Number of instances described in this chunk
Child Referents Array(Referent) Referents of child instances
Parent Referents Array(Referent) Referents of parent instances

The parent chunk (PRNT) defines the hierarchy relationship between every instance in the file.

There should be exactly one PRNT chunk.

Version field should currently always be zero.

Instance Count should be equal to the number of instances in the file header chunk, since each object should have a parent.

Child Referents and Parent Referents should both have length equal to Instance Count. The parent of the ID at position N in Child Referents is a child of the ID at position N in Parent Referents.

A null parent referent (-1) indicates that the object is a root instance. In a place, that means the object is a child of DataModel. In a model, that means the object should be placed directly under the object the model is being inserted into.

END Chunk

The END chunk has this layout:

Field Name Format Value
Magic Value 9 bytes Always </roblox>

The ending chunk (END) signifies the end of the file.

The END chunk must not be compressed. It is used as a rough form of file validation when uploading places to the Roblox website.

Data Types

String

Type ID 0x01

The String type is stored as a length-prefixed sequence of bytes. The length is stored as an untransformed 32-bit integer. String values are UTF-8 encoded.

Field Name Format Value
Length u32 The length of the string
Data Array(Bytes) The actual bytes that make up the string

When an array of String values is present, they are stored in sequence without any modification.

Bool

Type ID 0x02

The Bool type is stored as a single byte. If the byte is 0x00, the bool is false. If it is 0x01, it is true.

When an array of Bool values is present, they are stored in sequence.

Int32

Type ID 0x03

The Int32 type is stored as a big-endian transformed 32-bit integer.

When an array of Int32 values is present, the bytes of the integers are subject to byte interleaving.

Float32

Type ID 0x04

The Float32 type is stored using the Roblox float format and is little-endian. This datatype is also called float or single.

When an array of Float32 values is present, the bytes of the floats are subject to byte interleaving.

Float64

Type ID 0x05

The Float64 type is stored using the IEEE-754 format and is little-endian. This datatype is also called double.

When an array of Float64 values is present, they are in sequence with no transformations.

UDim

Type ID 0x06

The UDim type is stored as a struct composed of a Float32 and an Int32:

Field Name Format Value
Scale Float32 The Scale component of the UDim
Offset Int32 The Offset component of the UDim

When an array of UDim values is present, the bytes of each individual components are stored as arrays, meaning their bytes are subject to byte interleaving.

Two UDim values {1, 2} and {3, 4} look like this: 7f 80 00 80 00 00 00 00 00 00 00 00 00 00 04 08.

The first 8 bytes (7f 80 00 80 00 00 00 00) represent the Scale values of the UDim values. The latter 8 bytes (00 00 00 00 00 00 04 08) represent the Offset values. From there, the values are paired off, so that the first value in each array make up the components of the first UDim, and so on.

UDim2

Type ID 0x07

The UDim2 type is a struct composed of two UDim values, one for each axis:

Field Name Format Value
X UDim The X component of the UDim2
Y UDim The Y component of the UDim2

UDim2 is stored as four arrays of component values in the order X.Scale, Y.Scale, X.Offset, Y.Offset. Each array is separately byte interleaved.

An encoded UDim2 with value {0.75, -30, -1.5, 60} looks like this: 7e 80 00 00 7f 80 00 01 00 00 00 3b 00 00 00 78.

Ray

Type ID 0x08

The Ray type is a struct composed of six little-endian f32 values, making up the components of the Origin and then the Direction of the Ray:

Field Name Format Value
Origin X f32 The X component of the Origin of the Ray
Origin Y f32 The Y component of the Origin of the Ray
Origin Z f32 The Z component of the Origin of the Ray
Direction X f32 The X component of the Direction of the Ray
Direction Y f32 The Y component of the Direction of the Ray
Direction Z f32 The Z component of the Direction of the Ray

The components are stored in order without any additional transformations. When an array of Ray values is present, they’re stored in order but otherwise without transformation.

Faces

Type ID 0x09

The Faces type is a single byte used as a bit field. The low 6 bits represent the Front, Bottom, Left, Back, Top, and Right faces, in that order. The remaining two bits have no meaning. Faces is stored as an array of bytes with no transformations or interleaving.

Three encoded Faces with values Front, Back, Top and Bottom, Left, Right looks like this: 01 18 26.

Axes

Type ID 0x0a

The Axes type is a single byte used as a bit field. The low three bits represent the X, Y, and Z axes, in that order. The remaining five bits have no meaning. Axes is stored as an array of bytes with no transformations or interleaving.

Three encoded Axes with values X, X Y, and X Z look like this: 01 03 05.

BrickColor

Type ID 0x0b

The BrickColor type is a single untransformed big-endian u32 that represents the Number of a BrickColor. When an array of BrickColor values is present, the Numbers are byte interleaved but otherwise are unchanged.

As an example, three encoded BrickColor values Really red (1004), Bright green (37), and Really blue (1010) look like this: 00 00 00 00 00 00 03 00 03 EC 25 F2.

Color3

Type ID 0x0c

The Color3 type is a struct composed of three Float32 values:

Field Name Format Value
R Float32 The R component of the Color3
G Float32 The G component of the Color3
B Float32 The B component of the Color3

Color3 is stored as three arrays of components in the order R, G, B. Each array is separately byte interleaved.

An encoded Color3 with RGB value 255, 180, 20 looks like this: 7f 00 00 00 7e 69 69 6a 7b 41 41 42.

Vector2

Type ID 0x0d

The Vector2 type is a struct composed of two Float32 values:

Field Name Format Value
X Float32 The X component of the Vector2
Y Float32 The Y component of the Vector2

Vector2 is stored as two arrays of components in the order X, Y. Each array is separately byte interleaved.

Two encoded Vector2 values -100.80, 200.55, 200.55, -100.80 look like this: 85 86 93 91 33 19 35 9a 86 85 91 93 19 33 9a 35

Vector3

Type ID 0x0e

The Vector3 type is a struct composed of three Float32 values:

Field Name Format Value
X Float32 The X component of the Vector3
Y Float32 The Y component of the Vector3
Z Float32 The Z component of the Vector3

Vector3 is stored as three arrays of components in the order X, Y, Z. Each array is separately byte interleaved.

Two encoded Vector3 values 1, 2, 3 and -1, -2, -3 look like this: 7F 7F 00 00 00 00 00 01 80 80 00 00 00 00 00 01 80 80 80 80 00 00 00 01.

CFrame

Type ID 0x10

The CFrame type is more complicated than other types. To save space, there are 24 special cases where only the CFrame’s position is saved. The special case’s ID is written as a single byte.

If the byte is 00, a CFrame looks like this:

Field Name Format Value
ID u8 Always 00 in this case.
Orientation Array of 9 f32 values The rotation matrix of the CFrame. It represents the RightVector, UpVector, and LookVector, in that order.
Position Vector3 The position of the CFrame.

In this case, the Orientation field is stored as nine untransformed IEEE-754 standard 32-bit floats.

If the ID is not 00, it will be a value from the following table. In this case, the Orientation field isn’t present and is instead equivalent to the angles paired with the ID in the table. Rotations in this table are in degrees and are applied in the order Y -> X -> Z.

ID Rotation ID Rotation
02 (0, 0, 0) 14 (0, 180, 0)
03 (90, 0, 0) 15 (-90, -180, 0)
05 (0, 180, 180) 17 (0, 0, 180)
06 (-90, 0, 0) 18 (90, 180, 0)
07 (0, 180, 90) 19 (0, 0, -90)
09 (0, 90, 90) 1b (0, -90, -90)
0a (0, 0, 90) 1c (0, -180, -90)
0c (0, -90, 90) 1e (0, 90, -90)
0d (-90, -90, 0) 1f (90, 90, 0)
0e (0, -90, 0) 20 (0, 90, 0)
10 (90, -90, 0) 22 (-90, 90, 0)
11 (0, 90, 180) 23 (0, -90, 180)

When an array of CFrame values is present, for each value the ID is stored followed by the Rotation field if it’s present. Then, an array of Vector3 values that represent the Position field of each CFrame.

Two CFrame values with the components CFrame.new(1, 2, 3) and CFrame.new(4, 5, 6)*CFrame.Angles(7, 8, 9) are stored as 02 00 4B C0 07 3E 08 9C 75 3D 95 46 7D 3F 1D 25 90 BE 58 6C 74 BF 84 C5 C3 3D 1E 4A 73 3F 6F 19 95 BE 9F A6 E0 BD 7F 81 00 00 00 00 00 00 80 7F 00 22 00 D4 00 B2 80 81 80 80 00 00 00 00.

The first part (the ID and Rotation array) is: 02 00 4B C0 07 3E 08 9C 75 3D 95 46 7D 3F 1D 25 90 BE 58 6C 74 BF 84 C5 C3 3D 1E 4A 73 3F 6F 19 95 BE 9F A6 E0 BD, which is an split into 02 and 00 4B C0 07 3E 08 9C 75 3D 95 46 7D 3F 1D 25 90 BE 58 6C 74 BF 84 C5 C3 3D 1E 4A 73 3F 6F 19 95 BE 9F A6 E0 BD.

The second part (the Position array) is: 7F 81 00 00 00 00 00 00 80 7F 00 22 00 D4 00 B2 80 81 80 80 00 00 00 00.

Enum

Type ID 0x12

The Enum type is an unsigned 32-bit integer. It is stored as big endian and is subject to byte interleaving.

Referent

Type ID 0x13 The Referent type represents a specific Instance in the file and is stored as an Int32. After untransforming a Referent, a value of -1 represents the so-called ‘null referent’. In a PROP chunk, a null referent represents a property with no set value: for example, the default value of ObjectValue.Value.

An array of Referent values is stored as an array of Int32 values, and as a result they are subject to byte interleaving. When reading an array of Referent values, they must be read accumulatively. That is to say that the ‘actual’ value of the Referent is the value of the read value plus the preceding one.

Without accumulation, Referent values read from a file may look like this. This is incorrect:

Referent 1 Referent 2 Referent 3 Referent 4 Referent 5 Referent 6
1619 1 4 2 3 5

The correct interpretation of this data, with accumulation, is:

Referent 1 Referent 2 Referent 3 Referent 4 Referent 5 Referent 6
1619 1620 1624 1626 1629 1634

Vector3int16

Type ID 0x14

The Vector3int16 type is stored as three little-endian i16 values:

Field Name Format Value
X i16 The X component of the Vector3int16
Y i16 The Y component of the Vector3int16
Z i16 The Z component of the Vector3int16

Multiple Vector3int16 values are stored in sequence without any transformations or interleaving. Two Vector3int16 values 1, 2, 3 and -1, -2, -3 are stored like this: 00 01 00 02 00 03 FF FF FE FF FD FF.

NumberSequence

Type ID 0x15

The NumberSequence type is stored as a u32 indicating how many NumberSequenceKeypoint values are in the sequence followed by an array of NumberSequenceKeypoint values:

Field Name Format Value
Keypoint count u32 The number of keypoints in the sequence
Keypoints Array(NumberSequenceKeypoint) The data for the keypoints

NumberSequenceKeypoint is a struct composed of the following fields:

Field Name Format Value
Time f32 The time for the keypoint
Value f32 The value of the keypoint
Envelope f32 The envelope for the keypoint

When multiple NumberSequence values are present, they are stored in sequence with no transformation or interleaving. Two NumberSequence values

NumberSequence.new(
	NumberSequenceKeypoint.new(0, 0),
	NumberSequenceKeypoint.new(0.5, 1),
	NumberSequenceKeypoint.new(1, 1, 0.5)
)
NumberSequence.new(
	NumberSequenceKeypoint.new(0, 1),
	NumberSequenceKeypoint.new(0.5, 0.5, 0.5),
	NumberSequenceKeypoint.new(1, 0.5)
)

look like this: 03 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 3f 00 00 80 3f 00 00 00 00 00 00 80 3f 00 00 80 3f 00 00 00 3f 03 00 00 00 00 00 00 00 00 00 80 3f 00 00 00 00 00 00 00 3f 00 00 00 3f 00 00 00 3f 00 00 80 3f 00 00 00 3f 00 00 00 00

ColorSequence

Type ID 0x16

The ColorSequence type is stored as a u32 indicating how many ColorSequenceKeypoint values are in the ColorSequence followed by an array of ColorSequenceKeypoint values:

Field Name Format Value
Keypoint count u32 The number of keypoints in the sequence
Keypoints Array(ColorSequenceKeypoint) The data for each keypoint

ColorSequenceKeypoint is a struct composed of the following fields:

Field Name Format Value
Time f32 The time value of the ColorSequenceKeypoint
R f32 The red component of the keypoint’s color value.
G f32 The green component of the keypoint’s color value.
B f32 The blue component of the keypoint’s color value.
Envelope (unused) f32 n/a; serialized, but not used

When multiple ColorSequence values are present, they are stored in sequence with no transformation or interleaving. Two ColorSequence values

ColorSequence.new(
	ColorSequenceKeypoint.new(0, Color3.FromRGB(255, 255, 255)),
	ColorSequenceKeypoint.new(0.5, Color3.FromRGB(0, 0, 0)),
	ColorSequenceKeypoint.new(1, Color3.FromRGB(255, 255, 255))
)
ColorSequence.new(
	ColorSequenceKeypoint.new(0, Color3.FromRGB(255, 0, 0)),
	ColorSequenceKeypoint.new(0.5, Color3.FromRGB(0, 255, 0))
	ColorSequenceKeypoint.new(1, Color3.FromRGB(0, 0, 255))
)

look like this: 03 00 00 00 00 00 00 00 00 00 80 3f 00 00 80 3f 00 00 80 3f 00 00 00 00 00 00 00 3f 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 80 3f 00 00 80 3f 00 00 80 3f 00 00 80 3f 00 00 00 00 03 00 00 00 00 00 00 00 00 00 80 3f 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 3f 00 00 00 00 00 00 80 3f 00 00 00 00 00 00 00 00 00 00 80 3f 00 00 00 00 00 00 00 00 00 00 80 3f 00 00 00 00.

NumberRange

Type ID 0x17

The NumberRange type is stored as two little-endian floats:

Field Name Format Value
Min f32 The minimum value of the range
Max f32 The maximum value of the range

Multiple NumberRange values are stored in sequence with no transformation or interleaving. Two NumberRange values NumberRange.new(0, 0.5) and NumberRange.new(0.5, 1) look like this: 00 00 00 00 00 00 00 3f 00 00 00 3f 00 00 80 3f.

Rect

Type ID 0x18

The Rect type is a struct composed of two Vector2 values:

Field Name Format Value
Min Vector2 The minimum value of the Rect
Max Vector2 The maximum value of the Rect

Rect is stored as four arrays of Float32s in the order Min.X, Min.Y, Max.X, Max.Y. Each array is subject to byte interleaving.

Two encoded Rect values with values Rect.new(-1, -10, 8, 9) and Rect.new(0, 1, 5, 6) look like this: 7f 00 00 00 00 00 01 00 82 7f 40 00 00 00 01 00 82 81 00 40 00 00 00 00 82 81 20 80 00 00 00 00.

PhysicalProperties

Type ID 0x19

The PhysicalProperties type contains a flag which may be followed by a CustomPhysicalProperties value. CustomPhysicalProperties is a struct composed of five f32 values:

Field Name Format Value
Density f32 The density set for the custom physical properties
Friction f32 The friction set for the custom physical properties
Elasticity f32 The elasticity set for the custom physical properties
FrictionWeight f32 The friction weight set for the custom physical properties
ElasticityWeight f32 The elasticity weight set for the custom physical properties

If there is no CustomPhysicalProperties value, a PhysicalProperties is stored as a single byte of value 0. Otherwise, it is stored as a byte of value 1 immediately followed by a CustomPhysicalProperties stored as little-endian floats (in the same order as the above table). When there are multiple PhysicalProperties present, they are stored in sequence with no transformations or interleaving.

A default PhysicalProperties (i.e. no custom properties set) followed by a PhysicalProperties of value PhysicalProperties.new(0.7, 0.3, 0.5, 1, 1) looks like this: 00 01 33 33 33 3f 9a 99 99 3e 00 00 00 3f 00 00 80 3f 00 00 80 3f.

Color3uint8

Type ID 0x1a

The Color3uint8 type is a struct made up of three bytes, one for each component:

Field Name Format Value
R u8 The R component of the Color3uint8
G u8 The G component of the Color3uint8
B u8 The B component of the Color3uint8

Color3uint8 is stored as three consecutive arrays of components in the order R, G, B. It is not subject to any transformation or byte interleaving.

Two Color3uint8 values with the values 0, 255, 255 and 63, 0, 127, respectively, look like this: 00 3f ff 00 ff 7f.

Int64

Type ID 0x1b

The Int64 type is stored as a big-endian transformed 64-bit integer.

When an array of Int64 values is present, the bytes of the integers are subject to byte interleaving.

SharedString

Type ID 0x1c

SharedString values are stored as an Interleaved Array of u32 values that represent indices in the SSTR string array.

OptionalCoordinateFrame

Type ID 0x1e

OptionalCoordinateFrame is stored the same way as CFrame, but with a couple interesting differences:

  • Immediately following the type ID for OptionalCoordinateFrame is the type ID for CFrame (10);
  • At the end of the chunk there is an array of Bool values (preceded by the respective type ID, 02) that indicates which OptionalCoordinateFrame values have a value.

An OptionalCoordinateFrame with value CFrame.new(0, 0, 1, 0, -1, 0, 1, 0, 0, 0, 0, 1) followed by an OptionalCoordinateFrame with no value looks like this: 10 0a 02 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 7f 00 00 00 00 00 00 00 02 01 00. Note that the valueless OptionalCoordinateFrame is written as the identity CFrame with its corresponding boolean 00 codifying its valuelessness.

Data Storage Notes

Integer Transformations

Some integers may be subject to a transformation to make them more compressable.

To transform an integer: if x greater than or equal to zero, transform it with 2 * x. Otherwise, use 2 * |x| - 1. In most compilers this is equivalent to (x << 1) ^ (x >> 31) for 32-bit integers. For 64-bit integers, the same format is used but with 63 instead of 31.

To untransform one: if x is divisible by 2, untransform it with x / 2. Otherwise, use -(x + 1) / 2. This is equivalent to (x >> 1) ^ -(x & 1).

Untransforming with bitwise operators requires casting to an unsigned integer in some cases because x >> 1 will result in a negative number if x is negative.

Byte Interleaving

When stored as arrays, some data types have their bytes interleaved to help with compression. Cases where byte interleaving is present are explicitly noted.

When the bytes of an array are interleaved, they’re stored with the first bytes all in sequence, then the second bytes, then the third, and so on. As an example, the sequence A0 A1 B0 B1 C0 C1 is stored as A0 B0 C0 A1 B1 C1.

Viewed another way, it means that the bytes are effectively stored in ‘columns’ rather than ‘rows’. If an array of four 32-bit integers were viewed as a 4x4 matrix, for example, it would normally look like this:

  Column 1 Column 2 Column 3 Column 4
Row 1 A0 A1 A2 A3
Row 2 B0 B1 B2 B3
Row 3 C0 C1 C2 C3
Row 4 D0 D1 D2 D3

When interleaved, the same array would instead look like this:

  Column 1 Column 2 Column 3 Column 4
Row 1 A0 B0 C0 D0
Row 2 A1 B1 C1 D1
Row 3 A2 B2 C2 D2
Row 4 A3 B3 C3 D3

Roblox Float Format

Some data types do not follow the IEEE-754 standard format for 32-bit floating point numbers. Instead, they use a proprietary format where the sign bit is after the mantissa.

Format Bit Layout
Standard seeeeeee emmmmmmm mmmmmmmm mmmmmmmm
Roblox eeeeeeee mmmmmmmm mmmmmmmm mmmmmmms

Where s is the sign bit, e is an exponent bit, and m is a mantissa bit.

As a practical example, below is a comparison of how -0.15625 is stored:

Format Binary View Byte View
Standard 10111110 00100000 00000000 00000000 be 20 00 00
Roblox 01111100 01000000 00000000 00000001 7c 40 00 01