dlasher/openNemo-Cascade-2-30B-A3B-ENG:IQ4

dlasher/ openNemo-Cascade-2-30B-A3B-ENG:IQ4_NL

21 Downloads Updated 2 weeks ago

pulled as BF16, vocab slice (combined_en_huge) requant to IQ4_NL (combined_en_huge) -- (English + Coding focus) - ~32% faster than unsloth Q4_K_M

tools thinking

ollama run dlasher/openNemo-Cascade-2-30B-A3B-ENG:IQ4_NL

curl http://localhost:11434/api/chat \
  -d '{
    "model": "dlasher/openNemo-Cascade-2-30B-A3B-ENG:IQ4_NL",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='dlasher/openNemo-Cascade-2-30B-A3B-ENG:IQ4_NL',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'dlasher/openNemo-Cascade-2-30B-A3B-ENG:IQ4_NL',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 2 weeks ago

2 weeks ago

a54695d06a99 · 18GB

model

archnemotron_h_moe

parameters31.2B

quantizationIQ4_NL

18GB

params

{ "min_p": 0.05, "num_batch": 2048, "num_ctx": 262144, "presence_penalty": 0, "r

242B

template

{{- if .System }}<|im_start|>system {{ .System }}<|im_end|> {{- end }} {{- range .Messages }} {{- if

286B

Readme

NOTE: This model has been heavily vocab-sliced to reduce non-coding and non-English tokens. YMMV. (Mandarin, Korean, Vietnamese, Hindi, Arabic, Japanese, Russian)

==================================================
 PERPLEXITY QUALITY SCORECARD
 Base Model PPL      : 6.1805 (16-bit reference)
 Quantized Model PPL : 6.1942 (IQ4_NL target)
 Perplexity Drift    : +0.0137 (+0.22%)
 Quality Assessment  : [PASS] EXCELLENT (Near-Lossless / Gold Standard)
==================================================

SOURCE: https://huggingface.co/empero-ai/openNemo-Cascade-2-30B-A3B

Pure-PyTorch drop-in replacement for NVIDIA’s Nemotron-Cascade-2-30B-A3B.

Removes all external CUDA kernel dependencies (mamba-ssm, causal-conv1d) and replaces them with native PyTorch operations, making the model fully compatible with bitsandbytes 4-bit quantization and QLoRA fine-tuning on consumer GPUs.

30B total parameters, 3B active per token. Loads in 17 GB VRAM with 4-bit quantization.