Giant open models now ship in 4-bit: what it means for memory planning

Eight models in our catalog ship their main checkpoint in a 4-bit-class format, and nineteen more add official 4-bit builds. See the file sizes and how they change the hardware you need.

  • Analysis
  • 4 min read
  • Published
  • Updated

For years, open models were released in 16-bit and the community compressed them afterwards. In our catalog that has changed for the largest models. Among current models with 25B or more parameters, eight publish their main checkpoint in a 4-bit-class format, and nineteen more publish an official 4-bit build too. If the terms are new, read quantization explained first.

Eight models whose main checkpoint is 4-bit-class

Model Parameters Format Checkpoint Our estimate
Kimi K3.1 Pro 2.8T MXFP4 1,561 GB about 1,717 GB
Kimi K3 2.8T MXFP4 1,561 GB about 1,717 GB
DeepSeek V4 Pro 1.6T FP4 experts, FP8 elsewhere 893 GB about 982 GB
Kimi K2.7 Code 1.03T INT4 595 GB about 655 GB
MiMo-V2.6 Pro 1.02T MXFP4 experts, FP8 and BF16 elsewhere 573 GB about 630 GB
DeepSeek V4.1 Flash 552B Mixed FP8 and FP4 510 GB about 561 GB
MiniMax M3.1 Flash 428B FP4 252 GB about 277 GB
gpt-oss-120b 117B MXFP4 experts 65 GB about 72 GB

Kimi, MiMo, gpt-oss and DeepSeek V4 land at about 0.56 to 0.58 bytes per parameter, a little above the 0.53 that pure MXFP4 would cost, and MiMo, gpt-oss and DeepSeek V4 say outright that only the expert weights are 4-bit. DeepSeek V4.1 Flash is the outlier at 0.92 bytes per parameter: its checkpoint mixes FP8 and FP4, so more of it stays in FP8.

The developers describe this as deliberate. Moonshot says Kimi K3.1 Pro was trained in MXFP4 from the start, OpenAI post-trained gpt-oss with MXFP4 expert weights, MiniMax says M3.1 Flash is trained to be quantization-friendly, and NVIDIA pretrained Nemotron 3 Ultra with an NVFP4 recipe.

Nineteen more with an official 4-bit build

Model Original 4-bit build Share
LongCat-2.5 3,562 GB (BF16) 1,336 GB (W4A4) 38%
Inkling 1,909 GB (BF16) 594 GB (NVFP4) 31%
Muse Spark Open 1,644 GB (BF16) 533 GB (NVFP4) 32%
K-EXAONE 2.0 1,499 GB (BF16) 526 GB (NVFP4) 35%
Nemotron 3 Ultra 1,121 GB (BF16) 352 GB (NVFP4) 31%
Mistral Large 4 1,061 GB 610 GB (NVFP4) 57%
Reflection Beam 1,003 GB (BF16) 294 GB (NVFP4) 29%
A.X K2 694 GB (FP8) 398 GB (NVFP4) 57%
Motif 3 630 GB (BF16) 187 GB (NVFP4) 30%
Command A+ 438 GB (BF16) 132 GB (W4A4) 30%
Ornith 1.5 122B 250 GB (BF16) 78 GB (NVFP4) 31%
Nemotron 3.5 Super VL 249 GB (BF16) 75 GB (NVFP4) 30%
Ripple-65B Preview 130 GB (BF16) 57 GB (NVFP4) 44%
MiniCPM-V 4.7 35B-A3B 71 GB (BF16) 23 GB (int4) 32%
Nemotron 3.5 Transcribe 68 GB (BF16) 22 GB (NVFP4) 32%
Nanbeige4.3-33B 66 GB (BF16) 23 GB (INT4) 34%
Nemotron 3.5 Lightning 66 GB (BF16) 22 GB (NVFP4) 33%
Granite 4.2 30B 59 GB (BF16) 17 GB (MXFP4) 29%
Raon-1 mini preview 33 GB (MXFP8) 18 GB (NVFP4) 54%

A 4-bit build is typically 29 to 38% of the 16-bit size. The share is 44% for Ripple-65B, whose experimental NVFP4 build quantizes only the MLP weights, and 54 to 57% for Raon-1 mini preview, Mistral Large 4 and A.X K2 because their originals are already smaller than 16-bit.

What changes when you plan hardware

  • Use the real file size. For shipped 4-bit weights our estimator uses the actual download plus 10% instead of a generic formula, which gives about 1.7 TB for Kimi K3.1 Pro. The memory guide covers the context cache.
  • There is little room to shrink further. Going below 4 bits costs noticeable quality, so a 2.8-trillion-parameter model still needs a multi-node cluster.
  • Official 4-bit builds change who can run a model. Meta describes Muse Spark Open on two Apple M5 Ultra Mac Studios with MLX, and K-EXAONE 2.0 falls from 1,499 GB to 526 GB.
  • Check who made the build. Nota AI’s INT4 build of Solar Open 2 is 143 GB, but Upstage did not make it, so we do not count it.
  • Check hardware support. NVFP4 builds are aimed at NVIDIA Blackwell GPUs, and every other format has its own engine requirements.

For where each model lands, open the models by VRAM list.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Kimi K3.1 Pro Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster
Kimi K3 Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster
DeepSeek V4 DeepSeek 284B – 1.6T 49B active 1M MIT (Permissive) 176 GB Big-memory desktop
Kimi K2.7 Code Moonshot AI 1.03T 32B active 256K Modified MIT (Conditional) 655 GB GPU server
MiMo-V2.6 Xiaomi 9.4B – 1.02T 42B active 1M MIT (Permissive) 6.49 GB Laptop or phone
DeepSeek V4.1 Flash DeepSeek 552B 16B active 1M MIT (Permissive) 561 GB GPU server
MiniMax M3.1 Flash MiniMax 428B 23B active 1M MiniMax Community (Conditional) 277 GB Big-memory desktop
gpt-oss OpenAI 21B – 117B 5.1B active 128K Apache 2.0 (Permissive) 15 GB Laptop or phone
LongCat-2.5 Meituan 1.6T 48B active 1M MIT (Permissive) 1.1 TB GPU server
Inkling Thinking Machines Lab 276B – 975B 41B active 1M Apache 2.0 (Permissive) 188 GB Big-memory desktop
Muse Spark Open Meta 816B 42B active 1M Muse Model License (Conditional) 563 GB GPU server
K-EXAONE 2.0 LG AI Research 750B 37B active 256K Apache 2.0 (Permissive) 518 GB GPU server
Nemotron 3 Ultra NVIDIA 550B 55B active 1M OpenMDW-1.1 (Permissive) 379 GB Big-memory desktop
Mistral Large 4 Mistral AI 1.05T 52B active 1M Mistral Open-Weight (Conditional) 671 GB GPU server
Reflection Beam Reflection AI 501B 23.4B active 1M Apache 2.0 (Permissive) 323 GB Big-memory desktop
A.X K2 SK Telecom 688B 33B active 256K Apache 2.0 (Permissive) 438 GB Big-memory desktop
Motif 3 Motif Technologies 314B 13.2B active 256K MIT (Permissive) 206 GB Big-memory desktop
Command A+ Cohere 218B 25B active 128K Apache 2.0 (Permissive) 145 GB Big-memory desktop
Ornith 1.5 DeepReinforce 9.65B – 403B 17B active 256K MIT (Permissive) 6.66 GB Laptop or phone
Nemotron 3.5 NVIDIA 8B – 120B 12B active 1M OpenMDW-1.1 (Permissive) 5.52 GB Laptop or phone
Ripple-65B Preview Ricursive Intelligence 64.8B 1M Apache 2.0 (Permissive) 45 GB Workstation
MiniCPM-V 4.7 OpenBMB 1.55B – 35.2B 3B active 256K Apache 2.0 (Permissive) 1.07 GB Laptop or phone
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU
Nanbeige4.3-33B Nanbeige LLM Lab 33.1B 512K Apache 2.0 (Permissive) 23 GB Gaming GPU
Nemotron 3.5 Lightning NVIDIA 31.6B 3B active 1M OpenMDW-1.1 (Permissive) 22 GB Gaming GPU
Granite 4.2 IBM 3.7B – 29.3B 128K Apache 2.0 (Permissive) 2.55 GB Laptop or phone
Raon-1 mini preview KRAFTON 28.8B 4.3B active 256K Apache 2.0 (Permissive) 18 GB Gaming GPU
Solar Open 2 Upstage 250B 15B active 1M Upstage Solar (Conditional) 157 GB Big-memory desktop