INTELLECT-3 — Quantized (compressed-tensors for vLLM, GLM-4.5-Air MoE finetune)

This repository provides quantized runtime builds of
PrimeIntellect/INTELLECT-3, repackaged for vLLM using the compressed-tensors format.

TL;DR

Quantized branch: W4A16 (INT4 weights / A16 activations) and W8A16 (INT8 weights / A16 activations) for vLLM via --quantization compressed-tensors.

Same calibration recipe as our recent cards: 512 chat samples at 2048 tokens max from neuralmagic/LLM_compression_calibration (rendered with the model’s chat template).

Weight-only AWQ, group size 128, symmetric quant, lm_head left in higher precision, exported with save_compressed=True.

Parent is a GLM-4.5-Air MoE finetune; notes below cover MoE-specific considerations.

Revisions & Branches

The main branch is a landing page (model card + links). Runnable artifacts live in per-quant branches.

main — placeholder / landing page
W4A16 — 4-bit weights / 16-bit activations (compressed-tensors)
W8A16 — 4-bit weights / 16-bit activations (compressed-tensors)

Quick links

What’s inside (per revision)

Sharded quantized weights (*.safetensors) + index (model.safetensors.index.json)
config.json with compressed-tensors metadata (weight_format, quantization, quantization_config, etc.)
Tokenizer artifacts (tokenizer.json, tokenizer.model, merges/vocab as applicable)
Optional: chat_template.jinja (inherits the finetune’s chat style)

Exact file lists may differ between branches — see Files and versions for each revision.

Quantization & calibration details (same script/recipe family as previous card)

Method / flow

llmcompressor oneshot pipeline with an AWQModifier (weight-only).

Targets / exclusions

Quantize Linear layers across the model (including MoE expert linear projections).
Ignore lm_head (kept in higher precision).

Weights / grouping

INT4 (num_bits=4, type="int", symmetric=True)
Strategy: "group" with group_size=128 (Marlin-friendly)
Activations are not quantized (runtime A16: BF16/FP16)

Calibration dataset & preprocessing

Dataset: neuralmagic/LLM_compression_calibration, split train
NUM_CALIBRATION_SAMPLES = 512 (random subset with fixed seed)
MAX_SEQUENCE_LENGTH = 2048
Each sample’s messages list is rendered via the model tokenizer’s
apply_chat_template(..., tokenize=False), then tokenized with:
- max_length=2048, truncation=True, padding=False, add_special_tokens=False

Compression call

oneshot(..., max_seq_length=2048, num_calibration_samples=512, tokenizer=tokenizer) on the preprocessed dataset

Export for vLLM

Saved with save_compressed=True so vLLM reads the compressed-tensors runtime layout directly

GLM-4.5-Air MoE notes

Mixture-of-Experts (MoE) means most transformer blocks host multiple expert FFNs with a router/gating network that activates a subset per token.
Quantization impact: AWQ weight-only quantization is applied to expert Linear layers as well as shared projections; the router (small linear(s)) is quantized like other Linear layers.
Serving tips (vLLM):
- Ensure your vLLM build supports MoE routing for the GLM-family architecture.
- Throughput depends on expert parallelism + tensor parallelism; scale --tensor-parallel-size to your GPUs and mind interconnect bandwidth.
- Token-wise active experts increase KV-cache and memory pressure slightly; keep --max-model-len aligned with hardware.

Context length

Calibration context: up to 2048 tokens per sample (as above).
Model context window: inherited from PrimeIntellect/INTELLECT-3; quantization does not change rope/position encodings—only the numeric representation of the weights.

Quickstart — vLLM (compressed-tensors)

Install vLLM (recent version recommended):

pip install vllm

Serve (adjust to your hardware):

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
vllm serve TheHouseOfTheDude/INTELLECT-3_Compressed-Tensors \
  --quantization compressed-tensors \
  --tensor-parallel-size 8 \
  --max-model-len 2048 \
  --gpu-memory-utilization 0.70 \
  --dtype bfloat16

Example Chat Completions:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "TheHouseOfTheDude/INTELLECT-3_Compressed-Tensors",
    "messages": [
      {"role":"system","content":"You are INTELLECT — helpful, precise, and safe."},
      {"role":"user","content":"Outline a plan for multi-document retrieval with MoE models."}
    ],
    "max_tokens": 512,
    "temperature": 0.7,
    "top_p": 0.95
  }'

Note: compressed-tensors is a vLLM runtime format. Loading directly with vanilla 🤗 Transformers is not supported.
For Transformers, use a compatible export (e.g., GPTQ/AWQ for Transformers) or the full-precision finetune.

Prompting / chat template

This package follows the finetuned parent’s chat conventions. If a chat_template.jinja is present, libraries that support apply_chat_template will automatically format messages.

Guidelines:

Keep the system message concise (behavior, tone, safety constraints).
Provide clear user instructions; for multi-step tasks, list steps explicitly.

Intended use & safety

This quantization:

Does not change underlying behavior or content tendencies.
Only changes weight storage for efficient inference.

Apply appropriate content filters / policies for your deployment context.

Lineage

Finetuned parent: https://huggingface.co/PrimeIntellect/INTELLECT-3
This repo: Quantized child of the finetune (compressed-tensors for vLLM)

Hardware tips

100B+-class MoE models benefit from multi-GPU tensor parallel; interconnect bandwidth matters (NVLink/IB).
Long contexts are KV-cache heavy — tune --max-model-len and batch size.
Prefer BF16 on GPUs with native support; otherwise FP16.
Consider CUDA Graphs if stable in your environment.

Changelog

v1 (current) — Initial compressed-tensors W4A16 quantization with 512-sample / 2048-token AWQ calibration; vLLM-ready packaging.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for TheHouseOfTheDude/INTELLECT-3_Compressed-Tensors

Base model

zai-org/GLM-4.5-Air-Base

Finetuned

PrimeIntellect/INTELLECT-3

Quantized

(18)

this model