What are open weights? 2-bit and 4-bit quantization explained

You open a model announcement and see “open weights” next to a file called UD-Q2_K_XL. Short answer: open weights means the developer published the model’s files, so anyone can download and run them. 2-bit means every number inside is stored with just two bits: the file gets far smaller and the model gets a little blurrier. If you would rather skip the download, you can call models through an API instead; the CheapTokens docs show the first request.

The example below is real. Unsloth’s model page for MiniMax-M3.1-Flash (undated; checked 27 October 2026) lists GGUF files from 129.4 GB to 259.1 GB for a model we covered in the M3.1 Flash launch post. Every size and memory figure here comes from that page.

The words you keep meeting#

Picture a mixing desk with 428 billion dials. Training nudges every dial until the output sounds right. The final position of all dials is the model’s weights. Publish that list and you have an open-weight model; keep it on your own servers and the only way in is an API.

Two traps for beginners. “Open-weight” is not automatically “free for any use”: the license decides, and Unsloth’s page does not state it, so read the model card (MiniMaxAI/MiniMax-M3.1-Flash on Hugging Face) before building anything commercial. And “free to download” is not “free to run”, which the next sections show.

Why one model is a 130–260 GB download#

Each of those 428 billion weights is a number, and numbers take space. The rule of thumb is count × bits ÷ 8 = bytes. Try it before reading on.

Now scale up. Even at 2 bits per weight, 428 × 2 ÷ 8 = 107 GB is the floor for the whole model. The real 2-bit file is 143.9 GB, a bit above the floor, so the bit number in a file name is a label rather than an exact count.

The “23B active” figure is about speed, not storage. Only a slice of the model works on each token, which makes it quick, but the slice changes from token to token, so all 428B weights have to be loaded and waiting.

What 2-bit and 4-bit actually do#

Think of a ruler. Bits decide how many marks it has: 2 bits give 4 marks, 4 bits give 16, 8 bits give 256, and the usual 16 bits give 65,536. Quantization rounds every weight to the nearest mark. Fewer marks means a smaller file and a larger rounding error on each number.

Here is Unsloth’s lineup for the model, with the memory (RAM plus VRAM) it needs at 16K context for one session. Sizes are decimal GB.

Reading a name: UD-Q2_K_XL is an Unsloth Dynamic build at roughly 2 bits; the digit is the level and the letters are variants. Do not do arithmetic with the digit. The 8-bit file is only about 1.8 times the 2-bit one (259.1 ÷ 143.9 ≈ 1.8), not 4 times. Read the size off the download page.

What you give up at 2 bits#

Compare it to saving a photo as a smaller JPEG: most of the picture survives, and the damage shows up in the fine detail. Unsloth reports that its 2-bit file (143.9 GB) reaches a 91.5% weighted agreement with the original model across seven categories: coding, math, chat, tools, multilingual, reasoning and agentic.

Read that number carefully. It is Unsloth’s own measure, and the page does not spell out the baseline or the weighting, so treat it as a rough guide and not an independent test. Agreement is also not accuracy: it says how closely the small file behaves like the big one, not how often either is right. Small differences can add up over a long chain of steps, such as an agent running twenty tool calls, so try a quant on your own task before you rely on it. Higher levels stay closer to the original and cost more memory.

Download or call an API?#

Say you want to try the model after reading the announcement. Pick the line that sounds like you.

For the third line, one request is enough to see how it feels. This one goes to DeepSeek-V4-Flash, which has a 380K context window and answers up to 32,768 tokens long, at one price at any hour of the day:

curl https://api.cheaptokens.dev/v1/chat/completions \
  -H "Authorization: Bearer $CHEAPTOKENS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "deepseek-ai/DeepSeek-V4-Flash-0731", "messages": [{"role": "user", "content": "In two sentences: what does open-weight mean?"}], "max_tokens": 200}'

The answer comes back in choices[0].message.content, and the usage field tells you how many tokens you paid for. Not sure what a token is? The plain-English explanation takes five minutes. MiniMax-M2.7 is available the same way, with a 180K context window and up to 8,192 tokens per answer. Both use an OpenAI/Anthropic-compatible API, and one balance covers every model.

What does that look like on a bill? Say a small support bot sends 2,000 requests a day, each with 1,500 tokens in and 500 out. One such request costs $0.00015 on MiniMax-M2.7. Here is the whole day for each model, next to OpenRouter:

In total, with input and output counted together, CheapTokens comes out cheaper than OpenRouter. Current prices for every model are on the pricing page.

If you do have the hardware and want to compare speed and running costs, Run DeepSeek locally or use an API goes through it. And expect these words again soon: the open weights for Kimi K3.1 Pro are due on 4 November, as our launch post explains.