LLM Parameters Explained: What 1B, 8B, 70B, and 1T Actually Get You

What Are LLM Parameters?

A large language model is a neural network — a mathematical system composed of layers of interconnected numerical operations. Each connection in the network has an associated weight: a numerical coefficient that determines how strongly one piece of information influences the next. These weights are the parameters.

During training on massive text datasets — hundreds of billions of words drawn from books, websites, scientific literature, and code repositories — the model adjusts these weights iteratively through a process called gradient descent. The goal is to minimize the difference between the model's predictions and the actual next tokens in the training data. After trillions of these adjustments across billions of training examples, the weights collectively encode the model's knowledge: factual recall, reasoning patterns, language structure, domain expertise, and the subtle associations between concepts that make language meaningful.

A 7-billion parameter model has 7,000,000,000 of these numerical values. A 70-billion parameter model has ten times as many. A 1-trillion parameter frontier cloud model has more than 140 times as many as the 7B tier.

One clarification on terminology: parameters in this sense are the model’s trained weights, not the runtime settings such as temperature, top-p and top-k that shape how a model samples its next token.

Why does this matter for enterprise architects? Because parameter count is the primary driver of three decisions: the quality ceiling of the model, the hardware required to run it, and where it can be physically deployed. Larger models require more memory to store (all those weights must fit somewhere) and more compute to run (each token generation involves multiplying through all the parameter matrices). Understanding the hardware implications of each parameter tier is the prerequisite for making infrastructure decisions that are neither over-engineered nor prematurely limiting.

The Parameter-to-Capability Cheat Sheet

The following table is the core reference for enterprise technology selection decisions involving local or on-premises LLM deployment. Quality percentages represent approximate capability for typical enterprise business tasks including document analysis in the 10–20 page range. All models perform competently on simple requests; the quality gap emerges on complex, multi-step analytical tasks.

Note: Quality percentages represent approximate guidance for complex business tasks with 10–20 page document inputs. All models perform comparably on simple 1–2 sentence requests. Hardware requirements are for full-precision (FP16) operation; 4-bit quantization approximately halves memory requirements with minimal quality loss for business tasks. Assess models against your specific use cases — benchmarks are approximations, not guarantees.

The 1B, 3B and 7–8B rows in that table are the tiers commonly described as small language models, and their memory figures are what decide which laptops and AI PCs can run them without a discrete GPU.

The Largest LLMs by Parameter Count

The table below ranks the models whose parameter counts their developing labs published, largest first — from Meta’s unreleased Llama 4 Behemoth and the BAAI (Beijing Academy of Artificial Intelligence) and Google research systems down to the open-weights models an enterprise can actually download, including xAI’s Grok-1 and OpenAI’s gpt-oss-120b under Apache 2.0. Every figure is the one its lab documented. The frontier commercial assistants from OpenAI, Anthropic and Google do not disclose parameter counts for their flagship models, so any number circulating for those is an estimate rather than a published specification.

Total parameters versus active parameters

Every model at the top of that table is a mixture-of-experts design, and that changes what the headline number means. A dense model such as Llama 3.1 405B runs all 405 billion parameters for every token it generates. A sparse mixture-of-experts model such as DeepSeek-V3 stores 671 billion parameters but routes each token through roughly 37 billion of them. The stored total drives how much memory the weights occupy; the active count drives how much compute each token costs.

This is why “highest parameter count” and “most capable” have drifted apart. A 1-trillion-parameter sparse model can be cheaper per token than a 405-billion-parameter dense one, and a well-trained 70B dense model can outperform a larger sparse model on reasoning tasks. Compare the tier you are considering against the quality and hardware columns in the parameter-to-capability cheat sheet above rather than against the headline count.

What the largest models mean for an enterprise deployment

Only the open-weights models in that table can be deployed on-premises at all, and each one needs a multi-GPU cluster: Llama 3.1 405B runs on a single 8× H100 node at FP8, and the trillion-parameter sparse models need several such nodes. The research systems above them were never released, and none of this tier runs on a laptop. For most organizations its practical role is the cloud tier of a hybrid architecture: frontier capability for non-sensitive workloads, with a locally deployed 3B to 14B model handling the sensitive material that must never leave the building. For that split, see Hybrid AI Architecture, and for the local side of it, AirgapAI runs the 3B to 14B tier fully offline on current AI PC hardware.

At the opposite end of the range, the small-model tier has its own trade-off curve. For more information visit the SLM vs LLM comparison.

The 13x Compression Story: LLaMA 3’s 1B = LLaMA 2’s 13B

One of the most instructive data points for enterprise AI planning is the LLaMA 2-to-3 efficiency leap. Meta's LLaMA 3.2 1-billion parameter model achieves equivalent benchmark performance to LLaMA 2's 13-billion parameter model. The same quality. One thirteenth of the parameters. Roughly one thirteenth of the compute and memory requirements at inference time.

This 13x compression is not an anomaly. It is part of a consistent trajectory in open-source model development. As researchers apply better training data curation, improved architecture design, and more efficient attention mechanisms, the number of parameters required to reach a given quality level decreases with each major model generation. Microsoft's Phi-4 14B approaches GPT-4o Mini performance while running on workstation hardware. These gains compound.

For enterprise architecture planning, this trajectory has a direct implication: do not over-specify hardware for the model you need today. The device that runs a 3B model in 2026 will run the equivalent of a 2026 14B model within two model generations. Organizations that choose AI platforms supporting model flexibility — specifically platforms that allow model updates without application rebuilds — will continuously capture these efficiency gains. Platforms that let a team bring your own model keep that upgrade path open without an application rebuild. AirgapAI is designed with this model-agnostic architecture in mind, allowing organizations to upgrade the local model as the efficiency trajectory delivers better models in the same hardware envelope.

The 6–12 Month Lag Rule

The open source trajectory has established a consistent lag pattern: within 6 to 12 months of any given date, models matching the previous year's frontier capability typically become available for local deployment.

In practical terms: the models running locally on modern AI PCs today deliver quality equivalent to frontier cloud models from 12 months ago. Within the next 12 months, locally deployable models will reach the quality level that today's frontier cloud models provide. This is not a projection — it is an observed pattern that has held consistently across four generations of open-source model releases.

The 6–12 month lag rule has a critical strategic implication for organizations deliberating about local versus cloud deployment. If your primary objection to local LLM deployment is that current 7B–14B models are not quite as capable as frontier cloud models, that gap is closing on a predictable timeline. The organizations that build local AI capability now — deploying AirgapAI or equivalent local platforms on modern AI PC hardware — will be positioned to upgrade to equivalent-to-frontier capability within 12 months while retaining all the cost, data sovereignty, and operational benefits of local deployment.

The organizations waiting for local models to reach frontier quality before deploying will find that by the time that milestone arrives, a new frontier exists to chase. The correct entry point is now, with the current generation of local models, on a platform that supports model flexibility. See the broader framework for when to start versus when to wait in the Enterprise AI Strategy Guide and the AI Pilot Purgatory analysis.

“A 3B Model on a Laptop = Nov 2022 ChatGPT” — And Why That’s More Than Enough

The original ChatGPT released in November 2022 transformed how the world thought about AI. Within two months, it reached 100 million users. Businesses credited it with genuine productivity improvements. It was broadly considered the most capable AI assistant most people had ever interacted with.

A 3-billion parameter model running locally on a modern laptop — right now, today, disconnected from the internet — delivers approximately equivalent quality to that November 2022 ChatGPT.

That baseline is more than sufficient for the vast majority of enterprise knowledge work tasks: drafting emails, summarizing documents, answering policy questions, generating reports, translating text, reviewing contracts for standard provisions. The tasks that genuinely require frontier model quality — novel scientific reasoning, complex multi-step code generation with ambiguous requirements, multi-document synthesis of thousands of pages simultaneously — represent a small fraction of organizational AI usage.

Karpathy's observation points to something the parameter size debate often misses: the constraint on organizational AI value is rarely model quality. It is organizational readiness — the ability of employees to formulate clear prompts, integrate AI into their workflows, and act on AI outputs effectively. An organization that deploys a 3B model to 100% of its workforce and invests in AI Academy training will generate more business value than an organization that deploys a frontier model to 20% of its workforce and leaves the rest behind.

The 10-20-70 rule from The AI Strategy Blueprint reinforces this: 70% of AI success depends on people and process, not technology. A 3B local model with excellent user training delivers more value than a 1T cloud model with no training program. Read the full analysis at The 10-20-70 Rule for AI.