A tier-by-tier guide to picking local LLMs from 4GB to 128GB of memory, matching Gemma 4 E4B, Nanbeige 4.2, Bonsai 27B, Gemma 4 12B, Qwen 3.6-27B, ThinkingCap, Gemma 4T 31B, and Poolside's Laguna S 2.1 to real hardware. The core argument is that download size is not runtime memory, so the right model is the strongest one that fits comfortably with your actual context length, not the biggest one that barely loads.
Code Bear12 minTranscript found
Quick learning frame
Read this before watching.
Creative automation uses agents to accelerate production while keeping human taste in story, pacing, selection, and critique.
New playlist item from Code Bear; queued for transcript-backed review, topic mapping, and a practical learning artifact.
Skill you build: The ability to size a local model against your real VRAM budget by accounting for context, KV cache, and runtime overhead, then verify runtime support and workload fit before committing to a download.
Watch for the shift from claim to mechanism. The learning value is the point where the transcript reveals a repeatable action, tool boundary, context move, review habit, or artifact.
Concept diagram
Where this video fits.
01Brief
02Source
03Generation
04Selection
05Edit
06Taste Review
Deep lesson
Turn this video into working knowledge.
2,070 cleaned transcript words reviewed across 690 timed caption segments.
Thesis
Best Local AI Models for Every GPU (4GB to 128GB) teaches a practical creative automation move: A tier-by-tier guide to picking local LLMs from 4GB to 128GB of memory, matching Gemma 4 E4B, Nanbeige 4.2, Bonsai 27B, Gemma 4 12B, Qwen 3.6-27B, ThinkingCap, Gemma 4T 31B, and Poolside's Laguna S 2.1 to real hardware. The core argument is that download size is not runtime memory, so the right model is the strongest one that fits comfortably with your actual context length, not the biggest one that barely loads.
The goal is not to remember the video. The goal is to extract the operating principle, tie it to timestamped evidence, test how far the claim transfers, and make something reusable.
0:28
Download size lies
“actually fit before downloading it, and which impressive benchmark claims should be treated carefully. Because the file size shown on Hugging Face is not the model's real memory requirement. A model also needs memory for its context, temporary...”
Hugging Face file size excludes memory for context, temporary calculations, and the inference engine: the 1-bit Bonsai 27B downloads at roughly 3.9GB but measures about 5.2GB at a 4,000 token context and can exceed 11GB at 100,000 tokens. The model did not grow, its working memory did, so the right question is what runs comfortably for your workload rather than what barely fits. Take a model you already run locally and record its actual peak memory at a 4K context and again at your longest realistic prompt, then compare both numbers to the advertised download size.
6:13
Total versus active
“around 32 GB, Gemma 4T 31B becomes practical in a quantization-aware build. This gives you a strong general-purpose model with text, image, and audio support, a mature ecosystem, and enough capacity for demanding reasoning and coding. Gemma 4...”
Above 24GB the question shifts from whether you can run a capable model to how much quantization you still need: Gemma 4T 31B becomes practical in a quantization-aware build at around 32GB, while the Gemma 4 26B A4B mixture-of-experts stores roughly 25 billion parameters but activates only about 3.8 billion per token. Total parameters drive storage memory, active parameters drive per-token compute, so an MoE buys speed while every expert still has to be held in memory. Write a two-column table for three models you are considering listing total parameters versus active parameters, and note for each whether your bottleneck is memory capacity or generation speed.
7:46
Capacity is not bandwidth
“possible. One of the most interesting current options is Poolside's Laguna S 2.1. Laguna contains roughly 118 billion total parameters, but activates around 8 billion for each token. It is designed specifically for agentic coding and long-running software...”
At 128GB, Poolside's Laguna S 2.1 offers roughly 118 billion total parameters with about 8 billion active, a 1M-plus token context, and agentic coding results on Terminal Bench and SWE-Bench Pro, but its FP8 weights occupy about 121GB and leave nothing for the OS, engine, or context, so a lower-bit build is the practical choice. The same caution applies to Nvidia's DGX Spark: 128GB of unified memory answers whether a model loads, while much lower memory bandwidth answers how fast it will actually generate. For your own machine, write down its memory capacity and its memory bandwidth separately, then predict which of the two will limit you first on a long agentic coding run.
01
Brief
Start with this video's job: A tier-by-tier guide to picking local LLMs from 4GB to 128GB of memory, matching Gemma 4 E4B, Nanbeige 4.2, Bonsai 27B, Gemma 4 12B, Qwen 3.6-27B, ThinkingCap, Gemma 4T 31B, and Poolside's Laguna S 2.1 to real hardware. The core argument is that download size is not runtime memory, so the right model is the strongest one that fits comfortably with your actual context length, not the biggest one that barely loads. Treat "Brief" as the outcome you are trying to make visible, not a topic label. Anchor it to 0:28, where the video says: “actually fit before downloading it, and which impressive benchmark claims should be treated carefully. Because the file size shown on Hugging Face is not the model's real memory requirement. A model also needs memory for its context, temporary...”
02
Source
Use "Source" to locate the part of the creative automation workflow the video is demonstrating. Ask what changes in your real setup if this claim is true. Anchor it to 6:13, where the video says: “around 32 GB, Gemma 4T 31B becomes practical in a quantization-aware build. This gives you a strong general-purpose model with text, image, and audio support, a mature ecosystem, and enough capacity for demanding reasoning and coding. Gemma 4...”
03
Generation
Turn "Generation" into the reusable artifact for this lesson: A creative workflow board with critique criteria and review checkpoints. This is where watching becomes something you can inspect and reuse.
04
Selection
Use "Selection" as the application surface. Decide whether the idea touches a browser flow, a local file, a model choice, a source document, a UI, or a review step.
05
Edit
Use "Edit" to prove the lesson. The evidence should connect back to the video title, transcript anchors, and a concrete output, not a generic best-practice claim.
06
Taste Review
Use "Taste Review" to carry the idea forward: save the prompt, checklist, diagram, or operating rule that would make the next agent run better.
Example
Source-backed work packet
Convert the video into a scoped task that includes the transcript claim, target workflow, acceptance criteria, and proof. The output should be a creative workflow board with critique criteria and review checkpoints..
Example
Claim vs. demo brief
Separate what the speaker claims, what the demo actually proves, and what still needs outside verification before you adopt the workflow.
Example
Teach-back module
Transform the lesson into a definition, a mechanism diagram, one misconception, one practice exercise, and a check-for-understanding question.
Do not learn it wrong
Treating the title as the lesson without checking what the transcript actually says.
Letting the prompt drift into generic advice that could apply to any video in the playlist.
Copying the tool setup without identifying the operating principle that transfers to your own stack.
Skipping the artifact, which means the learning never becomes operational or inspectable.
Do not count this as learned until these are true.
01
State the transcript-backed claim in your own words: A tier-by-tier guide to picking local LLMs from 4GB to 128GB of memory, matching Gemma 4 E4B, Nanbeige 4.2, Bonsai 27B, Gemma 4 12B, Qwen 3.6-27B, ThinkingCap, Gemma 4T 31B, and Poolside's Laguna S 2.1 to real hardware. The core argument is that download size is not runtime memory, so the right model is the strongest one that fits comfortably with your actual context length, not the biggest one that barely loads.
02
Explain the practical stakes without hype: New playlist item from Code Bear; queued for transcript-backed review, topic mapping, and a practical learning artifact.
03
Map the idea onto the Brief -> Source -> Generation -> Selection -> Edit -> Taste Review sequence and name the weakest link.
04
Produce the artifact and include the evidence that proves it: A creative workflow board with critique criteria and review checkpoints.
Put it into practice
Give this grounded prompt to Codex or Claude after watching.
You are helping me turn one specific YouTube video into real, durable learning.
Source video:
- Title: Best Local AI Models for Every GPU (4GB to 128GB)
- URL: https://www.youtube.com/watch?v=T7WF9okSfrw
- Topic: Creative Automation
- My current learning frame: Pick the tier that matches your hardware, download the recommended safe model and one experimental alternative, then benchmark both on a real debugging task, a long document, one tool call, and one image while logging peak memory at your true context length.
- Why this matters: New playlist item from Code Bear; queued for transcript-backed review, topic mapping, and a practical learning artifact.
Transcript anchors from this exact video:
- 0:28 / Evidence 1: "actually fit before downloading it, and which impressive benchmark claims should be treated carefully. Because the file size shown on Hugging Face is not the model's real memory requirement. A model also needs memory for its context, temporary..."
- 2:46 / Evidence 2: "compression result. But compression is not free. The creators report that the 1-bit model retains around 89.5% of the original model's aggregate benchmark performance. They also acknowledge that long horizon agentic coding is not currently its strongest use..."
- 4:32 / Evidence 3: "experimenting. 24 GB. 24 GB is currently the consumer sweet spot. This is enough memory to run a strong 27 billion parameter model at a sensible 4-bit quantization instead of crushing it down to one or two bits."
- 6:13 / Evidence 4: "around 32 GB, Gemma 4T 31B becomes practical in a quantization-aware build. This gives you a strong general-purpose model with text, image, and audio support, a mature ecosystem, and enough capacity for demanding reasoning and coding. Gemma 4..."
- 7:46 / Evidence 5: "possible. One of the most interesting current options is Poolside's Laguna S 2.1. Laguna contains roughly 118 billion total parameters, but activates around 8 billion for each token. It is designed specifically for agentic coding and long-running software..."
- 10:26 / Evidence 6: "inside your own tools before trusting the benchmark numbers. For a 128 GB coding workstation, run a properly quantized Laguna build instead of attempting to fill every available byte with FP8 weights. Then decide how much context you..."
Your task:
1. Use the transcript anchors above as the primary source packet. If you add outside context, label it clearly as outside context and keep it secondary.
2. Create a source-check table with columns: timestamp, claim, what the demo proves, confidence, and what still needs verification.
3. Extract the actual teachable claims from the video. Do not invent claims that are not supported by the title, lesson frame, or transcript anchors.
4. Build a reusable learning artifact: A creative workflow board with critique criteria and review checkpoints.
5. Include:
- a plain-English definition of the core idea
- a diagram or structured model using this sequence: Brief -> Source -> Generation -> Selection -> Edit -> Taste Review
- 3 concrete examples that apply the video idea to real agentic work
- 2 failure modes the video helps prevent
- a checklist I can use the next time I run Codex or Claude
- one practical exercise with a clear done signal
6. Add a "learning transfer" section: what changes in my workflow tomorrow if I actually learned this?
7. Add a "source check" section that cites which transcript anchor supports each major takeaway.
Quality bar:
- Make this specific to "Best Local AI Models for Every GPU (4GB to 128GB)", not a generic Creative Automation essay.
- Prefer operational examples, failure modes, and reusable artifacts over broad definitions.
- Call out uncertainty instead of smoothing over weak evidence.
- If evidence is weak, say what transcript segment or timestamp needs review instead of guessing.
- Finish with a concise artifact I could paste into my learning app.
Misconceptions
What to stop believing.
Creative AI removes the need for taste.
It increases the need for taste because output volume explodes.
The best prompt is enough.
References, critique, iteration, and post-production matter just as much.
Practice studio
Learning only counts when you make something.
01
Transcript evidence map
Separate what the video actually says from what you already believe about the topic.
3 source-backed takeaways with timestamps, confidence, and a transfer note.02
One useful artifact
Apply the video to a real workflow and produce a creative workflow board with critique criteria and review checkpoints..
A reusable artifact with a done signal and one verification step.03
Teach-back card
Explain the lesson to someone who has not watched the video yet.
A 90-second explanation, one diagram, one example, and one misconception to avoid.
Recall check
Answer first, then reveal — without rewatching.
Why can a 3.9GB model download fail to run on a 4GB GPU?
In a mixture-of-experts model, what do total parameters versus active parameters each determine?
Why is a full FP8 build of Laguna S 2.1 a poor fit for a 128GB workstation?
Source shelf
Use the video as a doorway, then verify with primary sources.
