Local AI, without the jargon

Local AI glossary

Start with the simple explanation, see a concrete example, then go deeper when you want the technical definition. Every term also explains why it matters when you run a model on a phone.

35 terms, explained from simple to technical

0–9

1-bit quantization

Also: 1-bit model · one-bit model

A huge grid of full-precision weights funnels down into tiny one-bit cells holding only plus one or minus one27B model · too big+1-1-1+1-1+1≈1 bit each · ≈4 GB

A very aggressive way to shrink a model so that it uses much less storage and memory.

Example

A 27B model that would normally be far too large for a phone may become small enough to try when released in a purpose-built 1-bit format.

Technical explanation

A quantization scheme that represents each model weight with roughly one bit on average. Real file sizes are usually larger than the theoretical minimum because formats also store scales, metadata, and sometimes higher-precision values.

Why it matters on your phone

It can make unusually large models fit on phones, but compatibility, speed, and output quality depend heavily on the specific format and runtime.

4K context

Also: 4K context window · 4096-token context

Capacity gauge to a 4096-token limit: a chat fits inside, a long document overflows past the markChatDocument4096

The model can work with roughly 4,000 pieces of text from the current conversation at once.

Example

A short chat or article may fit inside 4K context. A long document may not, or the oldest parts of the conversation may need to be removed.

Technical explanation

A 4K context window usually means a limit near 4,096 tokens, although the exact usable amount and handling of input versus output vary by app and model.

Why it matters on your phone

Longer context consumes more memory through the KV cache, so AICanRun uses 4K as a consistent baseline for compatibility estimates.

7B model

Also: 7 billion parameter model · 7B

Three circles sized in proportion to 1B, 7B, and 70B parameters: B counts parameters, not gigabytesB = parameters, not GB1B7B70B

A model with about seven billion adjustable values inside it. The number describes model size, not its download size.

Example

The same 7B model might be around 4–5 GB in a Q4 version and much larger in Q8.

Technical explanation

The B suffix means billion parameters. Parameter count is a rough measure of model scale, but architecture, training, and quantization also strongly affect capability and memory use.

Why it matters on your phone

More parameters usually require a larger model file and more memory, but the quantization determines how many bytes those parameters occupy.

A

Active parameters

A large grid of parameters where only a small lit block is read for the current tokenTokenTotal 30BActive 3B

The part of a model that is actually used for one piece of text. Some models keep most of their parameters idle at each step.

Example

A mixture-of-experts model may contain 30B total parameters but use only 3B of them to generate each token.

Technical explanation

Active parameters are the subset of weights evaluated for a particular token or forward pass, most commonly discussed with sparse mixture-of-experts architectures.

Why it matters on your phone

Active parameter count can affect generation speed, while the full model file still has to fit in storage and usually in addressable memory.

C

Context window

Also: context length

A long row of token chips; a fixed window frame covers only the most recent chips the model can seeModel seesForgottenNewest

How much of your current conversation or document the model can keep in mind at once.

Example

Increasing context from 4K to 8K can let the model handle more text, but it also increases the memory used for conversation history.

Technical explanation

The context window is the maximum sequence length, measured in tokens, that a model and runtime can process in one inference session.

Why it matters on your phone

A longer context creates a larger KV cache and can turn a model that fits comfortably into one that runs out of memory.

CPU

Central Processing Unit

Also: central processing unit

A CPU die containing four large cores: few strong general-purpose workers, one shown activeCPUFew strong cores

The phone's general-purpose processor. It handles everyday work and can also run parts—or all—of a local AI model.

Example

A local AI app may use the CPU for model operations that are not supported or efficient on the phone's GPU.

Technical explanation

The central processing unit executes general-purpose instructions. Mobile inference runtimes use optimized vector instructions and multiple CPU cores for tensor operations.

Why it matters on your phone

CPU design affects speed and power use, but large-model generation on phones is often limited by memory bandwidth rather than raw CPU peak performance.

G

Generation speed

Also: decode speed · output speed

Two rows of words appearing over the same time span: a fast model emits many, a slow model fewFastSlowSame time

How quickly the model writes its answer after it has started responding.

Example

At 8 tokens per second, a roughly 300-word answer may take about 50 seconds to finish.

Technical explanation

Generation or decode speed is the autoregressive output rate, commonly reported in tokens per second. Each new token is generated using the tokens that came before it.

Why it matters on your phone

It describes the waiting experience after the first token, but it does not include model loading or prompt-processing delay.

GGUF

GGML Universal File — GGML is the library that runs the math, named after its author Georgi Gerganov (GG) + ML

One GGUF file with labeled compartments inside: a large weights block plus small metadata and tokenizer blocks.GGUFWeightsMetadataTokenizer

A common model-file format used by apps that run AI locally.

Example

When PocketPal asks you to choose a model file, you may see names ending in .gguf, with separate files for Q4, Q8, or other versions.

Technical explanation

GGUF is a binary format that stores model tensors and metadata for inference runtimes in the llama.cpp ecosystem. GGML is the underlying machine-learning tensor library (its author's initials plus ML); GGUF is that library's model file format. It supports multiple architectures and quantization types.

Why it matters on your phone

The GGUF file and quantization you download determine storage use, model-memory use, and whether your app can load the model.

GPU

Graphics Processing Unit

Also: graphics processing unit

A GPU die packed with a grid of 48 tiny cores; one whole row runs the same operation in parallelGPUSame opMany small cores

A processor built to do many similar calculations at the same time. AI apps can use it to speed up model work.

Example

An app may offload model layers to the phone's GPU instead of calculating every layer on the CPU.

Technical explanation

The graphics processing unit is a highly parallel processor. Inference runtimes can use compute APIs such as Metal or Vulkan to execute supported tensor operations on it.

Why it matters on your phone

GPU support can improve performance, but real speed also depends on the runtime, model format, memory bandwidth, and thermal limits.

I

Inference

Training builds the model once and is expensive; inference is using the finished model for every answerTraining · onceInference · every answer

Using a trained model to produce an answer. It is the running part, not the training part.

Example

When you ask a downloaded model a question in PocketPal and it replies on your phone, that is local inference.

Technical explanation

Inference is the forward computation performed by a trained model to produce predictions or output tokens from an input prompt.

Why it matters on your phone

AICanRun estimates the memory and generation speed of inference on a phone; it does not estimate training a model on that phone.

K

KV cache

Key–Value cache

Also: key-value cache

Each token of the reply adds one entry to the KV cache, so the cache bar grows in step with the replyReplyKV cacheOne entry per token · longer reply = more RAM

The model's working memory for the conversation so it does not have to reread everything from scratch for every new token.

Example

A longer conversation produces a larger KV cache. Doubling the context setting can roughly double this part of memory use for the same model and configuration.

Technical explanation

The key-value cache stores attention keys and values from earlier tokens during autoregressive inference. Its size depends on context length, architecture, precision, and batch settings.

Why it matters on your phone

The KV cache uses RAM in addition to the model file, which is why download size alone cannot tell you whether a model fits.

L

Large language model

Also: LLM

Text goes into a stack of model layers and comes out as ranked guesses for the next wordThe catsat on thePromptNext wordmatrugsofa

An AI model trained to understand and generate text.

Example

Llama, Gemma, and Qwen are families of language models. Some versions are small enough to run directly on a phone.

Technical explanation

A large language model is a neural network trained on large text datasets to predict tokens, enabling tasks such as generation, summarization, extraction, and question answering.

Why it matters on your phone

Model family alone does not tell you whether it fits; you also need the parameter count, model file, quantization, and available memory.

llama.cpp

The llama.cpp engine loads a GGUF file, powers apps, and runs on phones, laptops, and desktopsGGUFAppsllama.cpp

Popular software that lets many language models run locally on computers and phones.

Example

A phone app can use llama.cpp behind its interface to load a GGUF model and generate replies without sending the prompt to a server.

Technical explanation

llama.cpp is an open-source C/C++ inference runtime with CPU and accelerator backends, broad model-architecture support, and GGUF model loading.

Why it matters on your phone

Runtime support determines which files and quantizations an app can open and which parts of the phone's hardware it can use.

M

Memory headroom

Also: headroom

Usable memory bar mostly filled by the model; the small remaining gap is the headroomModel 6.5 GBLeftHeadroom

Memory left over after the model and app have taken what they need.

Example

If a model needs 6.5 GB and the phone has about 9 GB available, the estimated headroom is 2.5 GB.

Technical explanation

Memory headroom is the difference between estimated usable memory and the combined model weights, KV cache, and runtime overhead.

Why it matters on your phone

More headroom gives Android or iOS and other apps room to work. Very low headroom increases the chance of failed loading or the app being closed.

Memory bandwidth

A wide pipe streams many data blocks per second from memory to the chip; a narrow pipe streams fewMemoryMore GB/sFewer GB/s

How quickly the phone can move data between memory and its processors.

Example

If two phones can both fit the same model, the phone with higher usable memory bandwidth may generate tokens faster.

Technical explanation

Memory bandwidth is the rate at which data can be read from or written to memory, commonly expressed in gigabytes per second. Large-model decoding often repeatedly reads model weights and can be bandwidth-bound.

Why it matters on your phone

AICanRun uses memory bandwidth as one input to speed estimates, then labels those results as estimates unless a real measurement exists.

Memory mapping

Also: mmap · memory-mapped file

A model file on storage divided into pages; dashed arrows pull only the needed pages up into RAMRAMOn demandModel file on storage

A way for an app to let the operating system load pieces of a model file as they are needed.

Example

Instead of copying an entire GGUF file into a separate buffer at once, a runtime can map the file and let the operating system manage its pages.

Technical explanation

Memory mapping associates a file with a process's virtual address space. File-backed pages are brought into physical memory on demand and managed by the operating system.

Why it matters on your phone

Mapping reduces unnecessary copying, but the model still needs substantial address space and physical memory while it is being used.

Metal

An app reaches the Apple GPU through the Metal API: the bridge layer in between is highlightedAppMetalAPIGPU

Apple's system for letting apps use the GPU efficiently.

Example

An iPhone inference app may enable a Metal backend so supported model operations run on the Apple GPU instead of only on the CPU.

Technical explanation

Metal is Apple's low-overhead graphics and compute API for iPhone, iPad, and Mac hardware. Inference runtimes can provide Metal kernels for supported tensor operations.

Why it matters on your phone

Metal support can improve local inference performance, but it does not remove the model's memory requirements or iOS app-memory limits.

Mixture of experts

Also: MoE

A router sends each token to only two of eight expert blocks, whose outputs merge into the answerTokenRouter2 of 8

A model made of several specialist sections that only turns on some of them for each piece of text.

Example

A model may contain 30B parameters in total but activate only 3B parameters for each token.

Technical explanation

A mixture-of-experts model uses a router to select a sparse subset of expert feed-forward networks for each token, reducing computation relative to activating every parameter.

Why it matters on your phone

MoE models can be faster than their total size suggests, but the complete set of experts still makes the model file and memory footprint large.

Model file

A downloaded model file lands inside a bigger dashed RAM box that also needs room for runtime and chat memory4 GBNeeded in RAMFile+Runtime+Chat

The downloaded file that contains the model's learned data and the information an app needs to open it.

Example

A repository may offer several GGUF model files for the same model, such as Q4 and Q8 versions with different sizes.

Technical explanation

An inference model file stores tensors, architecture metadata, tokenizer information, and format-specific data required by a compatible runtime.

Why it matters on your phone

Its download size is the starting point for storage and memory planning, but runtime overhead and conversation memory must be added too.

Model weights

Also: weights

A model outline packed with a grid of tiny cells; a magnified callout shows one cell is just a learned number-0.4172×7BBillions of learned numbers

The billions of learned numbers inside a model that shape how it responds.

Example

Quantization stores the same model weights with fewer bits, making the model file smaller.

Technical explanation

Weights are learned numerical parameters in a neural network. During inference, matrix operations apply those values to the model's intermediate representations.

Why it matters on your phone

Weights usually account for most of a local model's file size and memory use.

N

NPU

Neural Processing Unit

Also: neural processing unit

An NPU die: a matrix of numbers flows through a multiply-accumulate block into an accumulator and outNPU×+ΣAI math only · efficient

A processor designed specifically for AI calculations.

Example

A phone may advertise a powerful NPU, but a GGUF app cannot benefit from it unless the app and model format support that hardware path.

Technical explanation

A neural processing unit is a specialized accelerator optimized for neural-network operations, often with vendor-specific supported data types, compilers, and execution frameworks.

Why it matters on your phone

NPU specifications alone do not predict local LLM speed. Runtime access and model compatibility matter just as much.

O

On-device AI

Also: local AI · on-device inference

A phone containing the model and the chat, with the link to a distant cloud cutModelOn deviceNo cloud

AI that runs on your own phone instead of sending the work to a remote server.

Example

After downloading a compatible model, you can ask it questions in airplane mode if the app itself does not require an online service.

Technical explanation

On-device AI executes model inference using local compute and memory resources. Implementations may use the CPU, GPU, NPU, or a combination of them.

Why it matters on your phone

It can work offline and keep prompts on the device, but it is limited by the phone's memory, speed, battery, thermals, app, and model.

P

Parameters

Also: parameter count

A magnifier over a model block reveals a grid of billions of tiny learned numbers7BModelLearned numbers

The adjustable numbers a model learned during training. Labels such as 3B or 27B tell you roughly how many it has.

Example

A 27B model has about 27 billion parameters and will usually require much more storage and memory than a 3B model at the same quantization.

Technical explanation

Parameters are learned tensors—primarily weights and biases—that determine a neural network's transformations. B denotes billions of parameters.

Why it matters on your phone

Parameter count helps compare model scale, but it must be combined with quantization and architecture to estimate file size and speed.

Prompt

System instructions, history, and the question bundled into one prompt flowing into the modelSystemHistoryQuestionPromptModel

The text or instructions you give the model before it answers.

Example

‘Summarize these meeting notes in three bullets’ is a prompt. The pasted meeting notes are also part of the prompt.

Technical explanation

A prompt is the input token sequence supplied to a generative model, often including system instructions, conversation history, user content, and formatting templates.

Why it matters on your phone

Longer prompts take more time to process and use more of the available context window.

Prompt processing

Also: prefill · prompt evaluation

Many prompt chips rush into the processor while the output line is still emptyPrompt inReadingNo output yet

The work the model does to read your prompt before it begins writing the answer.

Example

A long pasted document may cause a noticeable pause before the first word appears, even if the answer then streams quickly.

Technical explanation

Prompt processing, or prefill, evaluates the input sequence and populates the KV cache. Its throughput is distinct from autoregressive decode speed.

Why it matters on your phone

Tokens per second on AICanRun describes output generation; it does not fully describe the delay before the first token.

Q

Q1, Q2, Q4, and Q8

Also: Q4 · Q8 · quant levels · quant names

Four bars from Q1 to Q8 growing in file size, with a quality curve that climbs steeply then flattens: Q4 is the sweet spotQ1Q2Q4Q8Quality

Labels for differently compressed versions of a model. Lower numbers are usually smaller; higher numbers usually preserve more detail.

Example

A Q4 file may fit on a phone when Q8 does not. Q8 may preserve more of the original model but takes much more space.

Technical explanation

The Q number broadly indicates the target bits per weight, while suffixes such as K_M describe a particular quantization method or quality-size tradeoff. Exact formats and effective bit rates vary.

Why it matters on your phone

Choosing the right quant is often the difference between a model fitting comfortably, barely fitting, or not loading at all.

Quantization

Also: quant · quantized model

The same weight value stored in 16, 8, and 4 bits: identical squiggle, shrinking row of bit cells, shrinking file size16-bit · 13 GB8-bit · 7 GB4-bit · 4 GB

A way to make a model smaller so it needs less memory, usually with some tradeoff in output quality.

Example

An 8 GB Q8 version might be reduced to roughly 5 GB in Q4, allowing Q4 to fit on a phone where Q8 cannot.

Technical explanation

Quantization represents model values with lower numerical precision, such as 8-bit, 4-bit, or lower formats, often using per-block scales and format-specific encodings.

Why it matters on your phone

It directly changes download size, memory use, generation speed, and potentially output quality.

R

RAM

Random-Access Memory

Also: memory · random-access memory

Files sit on storage; loaded apps and a model occupy RAM, the working memory, while runningStorageRAMModel

The phone's short-term working space. Apps and models use it while they are running.

Example

A phone may have 12 GB of RAM, but Android or iOS and other processes need part of it before a model can load.

Technical explanation

Random-access memory is volatile working memory used for active code and data. It is different from persistent storage used for downloaded files and photos.

Why it matters on your phone

A model can be stored on a phone yet still fail to run if its weights, KV cache, and app need more RAM than the system makes available.

Runtime overhead

Stacked memory bar: model weights plus KV cache plus a small extra runtime slab often forgottenWeights 4 GBKV cache+ runtimeTotal > file size

Memory used by the app and model-running software in addition to the model itself.

Example

A 4 GB model file may need more than 4 GB of RAM after adding the inference engine, temporary buffers, and conversation memory.

Technical explanation

Runtime overhead includes executable code, allocator state, compute buffers, tensor metadata, temporary workspaces, and other implementation-specific memory.

Why it matters on your phone

AICanRun includes a runtime allowance so its fit calculation is not based on model-file size alone.

T

Thermal throttling

Over a long session temperature climbs to the thermal limit and tokens per second step down at that momentLimitTemptokens/s

The phone slowing itself down because it has become too hot.

Example

A model may generate quickly for the first few minutes and then slow down during a long session as the phone heats up.

Technical explanation

Thermal throttling reduces processor frequency, voltage, or permitted power when temperature or power limits are reached.

Why it matters on your phone

Short benchmarks may be faster than sustained use, so estimated speed should not be treated as a permanent guaranteed rate.

Time to first token

Also: TTFT · first-token latency

Timeline: load and prompt-processing segments precede the first token; output ticks follow evenly after itLoadProcessFirst tokenThe wait

How long you wait before the model begins showing its answer.

Example

An app may wait five seconds to load and read your prompt, then generate the rest of the reply at 10 tokens per second.

Technical explanation

Time to first token measures latency from request start until the first output token, including relevant loading, prompt processing, scheduling, and initial decode work.

Why it matters on your phone

A high output speed can still feel slow when model loading or prompt processing creates a long initial pause.

Token

The text RUNNING LOCAL AI split into four uneven token chips; one chip is only part of a wordRunning local AIRunninglocalAI~4 chars = 1 token

A small piece of text the model reads or writes. It may be a whole word, part of a word, punctuation, or code.

Example

A 300-word English answer is often around 400 tokens, but the ratio changes with language and content.

Technical explanation

A token is an integer unit produced by a model-specific tokenizer. The model processes sequences of token IDs rather than raw words or characters.

Why it matters on your phone

Context limits and generation speeds are measured in tokens, so tokens affect both memory use and waiting time.

Tokens per second

Also: tokens/s · tok/s · t/s

Nine token chips emitted inside a one-second bracket: the rate is 9 tokens per second, roughly 7 words9 tokens1 second≈ 7 words

How quickly a model writes its answer after it starts. A higher number usually means less waiting.

Example

At 8 tokens per second, a roughly 300-word answer may take about 50 seconds to finish.

Technical explanation

Tokens per second is autoregressive decode throughput: the number of output tokens generated per second after generation begins.

Why it matters on your phone

It helps compare how responsive model generation may feel, but it does not include all loading and first-token delays.

U

Unified memory

CPU, GPU, and NPU boxes all connect to one shared memory pool instead of separate poolsCPUGPUNPUOne shared memoryNo separate copies

Memory shared by the phone's CPU, GPU, and other processors instead of giving each one a completely separate pool.

Example

On an iPhone, the CPU and GPU can work with the same physical memory, but iOS still limits how much an individual app can safely use.

Technical explanation

A unified memory architecture exposes shared physical memory to multiple processors, reducing explicit copies while retaining operating-system and per-process allocation limits.

Why it matters on your phone

Unified memory helps local inference, but the phone's advertised RAM is still not the same as the amount available to one model app.

Usable RAM

Also: available memory · memory available for the model

A 12GB RAM bar: the system reserves a slice, only the rest is usable for a modelSystemUsable for model12 GB total~9 GB

The part of the phone's RAM that a model app can realistically use after the operating system takes its share.

Example

A 12 GB Android phone may have roughly 9 GB available for a large model in AICanRun's estimate, not the full advertised 12 GB.

Technical explanation

Usable RAM is an estimate of memory available to the inference process after system-resident usage, operating-system behavior, and practical app limits are accounted for.

Why it matters on your phone

AICanRun compares total model memory against usable RAM rather than advertised RAM to avoid an unrealistically optimistic fit result.