GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
esatapedico
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
esatapedico
Model Tree
Nanbeige/Nanbeige4.3-33B
Input Modalities
Output Modalities
Supported Functionality
gate_proj / up_proj / down_proj in all 52 layers — 156 tensors.nvfp4-pack-quantized), group size 16, FP8 E4M3 per-tensor weight scales, weight only (input_activations: null).q/k/v/o, lm_head, embeddings, and every norm. They keep the source dtype.requires_calibration_data: false).The exact recipe ships as recipe.yaml:
yaml
default_stage: default_modifiers: QuantizationModifier: targets: ['re:.*mlp\.(gate|up|down)_proj'] ignore: [lm_head, embed_tokens, self_attn, norm] scheme: NVFP4A16 bypass_divisibility_checks: false requires_calibration_data: false| File | Size | Notes |
|---|---|---|
model-00001-of-00002.safetensors | 19,935,445,572 bytes | shard 1 |
model-00002-of-00002.safetensors | 10,913,553,844 bytes | shard 2 |
model.safetensors.index.json | 67,595 bytes | shard map |
config.json | 10,582 bytes | resaved, carries quantization_config |
recipe.yaml | 268 bytes | the quantization recipe above |
generation_config.json | 10,582 bytes | carried from source |
tokenizer.json, tokenizer.model, tokenizer_config.json, special_tokens_map.json, added_tokens.json | — | source tokenizer and chat template |
Config facts: LlamaForCausalLM, model_type: llama, 52 layers, hidden 7168, 56 query / 8 KV heads, 512K max positions, vocab 166,912.
Nanbeige/Nanbeige4.3-33B, 14 shards, 66.3 GB, 471 tensors (52×9 + 3, no MTP), SHA-256 verified 14/14 against the HuggingFace LFS oids.weight_map before the run; dry run clean.pipeline='sequential', weight only, no calibration data.The GGUF ladder in esatapedico/Nanbeige4.3-33B-NVFP4-GGUF is converted from this checkpoint with convert_hf_to_gguf.py --outtype auto, then split into seven size tiers with per-tier override maps. This repository is the checkpoint those tiers came from; use it if you want to reproduce the conversion or re-quantize.
Derivative work built from Apache-2.0 artifacts. Credit belongs to:
"Nanbeige" is the name of the base model publisher. This repository is not affiliated with, sponsored by, or endorsed by the publisher.
Apache-2.0, identical to every upstream artifact (the base model ships Apache-2.0; no LICENSE file is bundled upstream, so this repo carries its own Apache-2.0 text with attribution). Quantization is a transformation, not new training. When redistributing, please retain attribution to the base model publisher as above.
I post updates on new conversions, benchmarks, and what I'm working on over on my site. Follow along there to keep up with new releases and the work in progress. If you'd like to support more of it, you will find the options there. I do this on consumer hardware and like seeing how far it goes. More is on the way.
🌐 esatapedico.com. Updates, work-in-progress, and ways to support.
This model card was written by an AI assistant at the request of the repository author, who did the engineering. As with any AI-generated text, there may be errors; please verify anything important (sizes, configs, commands) against the checkpoint files before relying on it.
default_stage: default_modifiers: QuantizationModifier: targets: ['re:.*mlp\.(gate|up|down)_proj'] ignore: [lm_head, embed_tokens, self_attn, norm] scheme: NVFP4A16 bypass_divisibility_checks: false requires_calibration_data: false