Sources for every figure, name and capability this video states or puts on screen.
Google publishes Gemma 3 in five sizes: 270M, 1B, 4B, 12B and 27B.
Context windows differ by size. The model card states: “Total input context of 128K tokens for the 4B, 12B, and 27B sizes, and 32K tokens for the 1B and 270M sizes.”
So the video’s “Gemma 3 has versions listed with a 128 thousand token context window” is correct as stated: it is the three larger sizes that carry 128K, not the whole family. The shot showing the context bar is labelled to that effect.
Qwen3-Coder-30B-A3B-Instruct is a mixture of experts model. The model card states “Number of Parameters: 30.5B in total and 3.3B activated”, which is the “thirty billion parameter version with only a fraction of the parameters active at once” the video describes.
Its context length is “262,144 natively”, extendable “up to 1M tokens using Yarn”, and the card describes it as “optimized for repository-scale understanding”. That supports both “advertises long context for repository scale work” and “some models talk about 256 thousand tokens or even more”: 262,144 is 256K.
The largest variant, Qwen3-Coder-480B-A35B-Instruct, is “480B in total and 35B activated”, which is the 480 billion parameter model named in the parameter ladder.
OpenAI released gpt-oss-120b and gpt-oss-20b as open weight models under Apache 2.0, described as “designed for powerful reasoning, agentic tasks, and versatile developer use cases”, with full chain of thought and configurable reasoning effort.
On the memory figure, the model card states the aim is to have “the gpt-oss-20b model run
within 16GB of memory”. That is the source for “designed to run on consumer style hardware
with around sixteen gigabytes of memory”.
gpt-oss-20b is 21B total parameters with 3.6B active, natively quantized to MXFP4. gpt-oss-120b is the larger of the pair and is the 120 billion parameter model named in the parameter ladder.
The project README supports every capability the video attributes to it: a “Plain C/C++ implementation without any dependencies” for CPU builds, “Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads GPUs via MUSA)” for GPU backends, and “CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity” for the hybrid case.
GGUF is llama.cpp’s model file format, and the repository carries the conversion tooling
(convert_hf_to_gguf.py) and the gguf-py package for it.
The CUDA line is also the basis for the video’s point about software support: the strength of the tooling around NVIDIA is a property of the tools, not of the silicon.
Idle TTL. LM Studio documents exactly the setting the video describes: “Idle TTL (technically: Time-To-Live) defines how long a model can stay loaded in memory without receiving any requests. When the TTL expires, the model is automatically unloaded from memory.” The default for JIT loaded models is 60 minutes, and the idle timer resets on each request. Auto Evict, alongside it, unloads a previously loaded model before loading a new one.
The estimator. lms load --estimate-only previews memory requirements before a model is
loaded. The estimate “accounts for factors like context length, flash attention, and whether
the model is vision-enabled”, and reports whether the model may be loaded given the
configured resource guardrails. That is the estimator the video recommends using instead of
assuming maximum GPU offload is best.
Apple silicon puts the CPU and GPU on one package over a single memory pool. Apple’s own developer material describes the consequence directly: with a unified memory architecture “the traditional management of copies between system RAM and video RAM goes away”, and Metal “exposes the UMA through shared resources that allow the GPU and CPU to read and write the same memory”, so resource management becomes synchronising access “rather than duplicating or shadowing data between system memory and video memory”.
Apple also states the practical consequence for this subject: the unified memory architecture “enables the entire chip to access a large single pool of memory”, which allows its machines “to run larger AI models completely on device”.
This is the whole basis for the video’s Apple Silicon chapter, and specifically for the claim that the difference is about avoiding a hard VRAM wall rather than about being faster.
The 70 billion parameter tier the video opens its parameter ladder with is a real and common open weight size; the 120 billion and 480 billion tiers are gpt-oss-120b and Qwen3-Coder-480B-A35B, both cited above. The 7 billion, 12 billion and 30 billion figures used in the quantization comparison are size classes rather than specific models, and the video treats them that way.
These are the video’s own judgements rather than sourceable facts, and the picture states them as opinion or does not put a number on screen for them at all.