AI Tools Field GuideReading a model's spec sheet
Four things in a model announcement matter to you: the generation, the tier, the context window, and whether it is on your platform. The rest is detail, including the parameter count, which is why a "35B" model is not a third as good as a "105B" one.
Why the spec sheet matters to a user
You will never choose a model by its parameter count. But announcements arrive monthly, colleagues forward them, and your data shop will talk about quantization and open weights when it proposes to host something locally.
Knowing the words lets you ask the right question: is it available on my platform, and is it better than what I have for the work I do?
Size and precision
| Term | What it means | What it means for you |
|---|---|---|
| 35B (parameters) | About 35 billion learned numbers: capacity, not speed or quality. | Small new models can beat big old ones; closed-weight sizes may be undisclosed. |
| Precision (16-bit, 8-bit, 4-bit) | Bits storing each parameter. | Lower precision saves memory and can speed inference; quality loss varies, especially at 4-bit on small/reasoning models. |
| Quantization (Q4, Q8, "quant") | Lowering precision after training: Q4 means four-bit. | Helps fit 35B on a workstation; hosted services manage it. |
| Open-weight / closed-weight | Published, downloadable parameters versus private weights accessed through a service. | Open weights can be hosted or run locally. Downloading authorizes nothing: official material needs command-accredited infrastructure. |
| Version (5.x, 6, 3.1) | The generation. | Compare reasoning capability; see Choosing a model. |
| Tier | Size-and-speed variants within a generation. | Fast for bulk, larger for hard work. "Pro" can mean size or reasoning depth. |
| Reasoning / thinking | Intermediate steps before answering; sometimes adjustable. | Slower; helps hard problems. See Thinking modes. |
| Multimodal | Reads images, documents, audio or video alongside text. | Enables scan/video uploads. |
| Benchmark | Score on a named test, often vendor-reported. | Not your task; seek independent reproduction. Fit and verification matter more. |
Parameters are the number people quote and the number that means least to a user. Capability comes from size, training data, training method and generation together, and the newer, smaller model routinely beats the older, larger one.
Precision and quantization matter when someone proposes to run a model on hardware they control. That is why open-weight models are quantized for local use, and why "we can run it on a workstation" usually means "at 4-bit."
| A 35B model at… | Memory needed | Quality |
|---|---|---|
| 16-bit | About 70 GB | Full |
| 8-bit | About 35 GB | Nearly full |
| 4-bit | About 20 GB | Small loss for drafting, larger for reasoning |
Open-weight means the parameters are published and anyone can download and run them. It does not mean the model is free of license terms, or that you may put official material through a copy on an unaccredited machine. The rules apply to the machine.
Context and cutoff
| Term | What it means | What it means for you |
|---|---|---|
| Context window | Maximum tokens attended to per call. | Shared with preamble, tools and files. See Tokens and the context window. |
| Knowledge cutoff | When training data ends. | Later facts need search/documents. Check the model card; self-reports can be guesses. |
A token is roughly three quarters of a word: 1,000 tokens ≈ 750 words. For every tool, each vendor documents the limit per model; assume tens of thousands of words, not unlimited.
The knowledge cutoff is the date after which the model knows nothing unless you tell it or it searches.
| Because of the cutoff… | So… |
|---|---|
| Last year's organization chart reads as current | Give it the current one, or let it search |
| "What is the latest version of this instruction?" gets a confident old answer | That is a question for a document tool |
| Asking the model its own cutoff gets a guess | The model card is the source |
Reading an announcement
Take a typical headline: a vendor releases "Model 6 Pro, one million tokens of context, with thinking, 12 points higher on a coding benchmark." Read it in four passes.
- Generation
- Tier
- Context
- Availability
The benchmark is the part to skip.
| Pass | Read for | In the example |
|---|---|---|
| Generation | Is it newer than what you have? Newer generations usually reason better. | Model 6 is newer than Model 5 |
| Tier | Which size is this, and is there a fast one for bulk work? | "Pro" is the large tier |
| Context | Does the window help with your long documents? | Large, subject to uneven attention |
| Availability | Is it in your product's model list yet? A commercial release and its rollout to your account are different events, sometimes weeks apart. | Not until your product shows it |
The benchmark was chosen by the vendor and measures a task that is not yours. The test that counts is a week of your own work with the source attached.
What to do with this
Next time an announcement lands in your inbox, write the four items in the margin and see how much of the rest was noise. Then read The harness is the product: the wrapper around the model shapes your experience more than the spec sheet does.