{ "format": "neural-download-model-family-v1", "id": "nemotron-cascade-2", "primary_packet_id": "rapid-nemotron-cascade-2-30b-a3b-q4km", "name": "Nemotron Cascade 2", "display_name": "Nemotron Cascade 2 · 30B-A3B", "summary": "NVIDIA's Nemotron Cascade 2, a 30B hybrid attention/Mamba mixture-of-experts with about 3B active per word. Runs on one Arc Pro B70 in 4-bit form.", "updated_at": "2026-08-24", "architecture": { "class": "Nemotron Cascade hybrid attention/Mamba MoE", "total_parameters_approx": 30000000000, "active_parameters_approx": 3000000000, "layers": 52, "experts": 128, "active_experts": 6, "shared_experts": 1, "kv_heads": 2, "evidence": "results/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km/README.md" }, "dimensions": { "weight_revision": ["nemotron-cascade-2-30b-a3b-931b595"], "weight_quantization": ["Q4_K_M"], "runtime": ["llama.cpp SYCL fdb1db877"], "tp": [1, 2, 4], "mtp": [0], "configured_max_context_tokens": [2048], "kv": ["f16"] }, "weight_revisions": [ {"id": "nemotron-cascade-2-30b-a3b-931b595", "label": "Nemotron-Cascade-2 30B-A3B", "role": "measured GGUF weights", "repository": "bartowski/nvidia_Nemotron-Cascade-2-30B-A3B-GGUF", "revision": "931b595fc71b7ca14fb9d935af011f69f7c0434c", "model_manifest": "results/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km/README.md"} ], "model_variants": [], "transfer_scope": { "status": "one measured Cascade 2 GGUF revision", "transfers": ["the stock llama.cpp hybrid/Mamba path and exact one-B70 launch controls"], "does_not_transfer": ["performance to Nemotron 3.5, another quantization, a vLLM Mamba cache mode, TP2/4, or longer context"], "evidence": "results/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km/README.md" }, "model_signals": { "b70_fit": {"band": "one-card measured", "scope": "Q4_K_M short-context rapid lane", "basis": "The verified 24,725,740,992-byte GGUF completed the strict suite on one B70.", "reviewed_at": "2026-08-24"}, "quality_evidence": {"band": "rapid snapshot only", "scope": "12 unique cache-zero prompts and clean visible output; no token IDs or broad model-quality evaluation", "evidence": ["data/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km-llamacpp-faon-cacheoff-ctx2048-confirm-realistic128-20260704T235714Z.json"]}, "popularity": {"state": "not-scored", "reason": "No dated popularity snapshot is stored."} }, "run_measurements": [ { "id": "nemotron-cascade-2-q4km-tp1-rapid", "state": "lab-measured", "revision": "nemotron-cascade-2-30b-a3b-931b595", "variant": "Q4_K_M", "quantization": "Q4_K_M", "runtime": "llama.cpp SYCL fdb1db877", "config": {"tp": 1, "mtp": 0, "kv": "f16", "configured_max_context_tokens": 2048}, "profile_id": "rapid-model-snapshots-b70-realistic-v1", "measurement_class": "strict rapid snapshot", "promotion_status": "promoted standalone row", "quality_scope": "12/12 cached_tokens=0 and realistic final gate passed; streamed text deltas, no token IDs", "workload": "12 unique cold prompts, 128 output tokens, median tokens 1-100 after TTFT, server and request prompt caches disabled", "metrics": {"decode_tok_s": [50.90422891211857], "ttft_ms": [449.1593260318041]}, "evidence": "data/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km-llamacpp-faon-cacheoff-ctx2048-confirm-realistic128-20260704T235714Z.json" } ], "series_measurements": [], "estimates": [], "packets": [ { "id": "rapid-nemotron-cascade-2-30b-a3b-q4km", "label": "Nemotron Cascade 2 30B-A3B Q4_K_M · TP1", "revision": "nemotron-cascade-2-30b-a3b-931b595", "quantization": "Q4_K_M", "runtime": "llama.cpp SYCL fdb1db877", "cards": 1, "status": "rapid research snapshot", "evidence_level": "B70-measured narrow baseline", "coverage": ["decode", "TTFT", "TP1", "cache-zero strict suite"], "grades": {"evidence": {"grade": "D", "scope": "exact Q4_K_M TP1 rapid row", "basis": "valid measured snapshot with a narrow workload and no clean-host or broad quality packet", "reviewed_at": "2026-08-24", "evidence": ["results/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km/README.md"]}}, "featured_metric": {"metric": "decode_tok_s", "measurement_id": "nemotron-cascade-2-q4km-tp1-rapid", "sample_index": 0, "value": 50.90422891211857, "unit": "tok/s", "workload": "12 unique cold prompts, 128 output tokens, median tokens 1-100 after TTFT, server and request prompt caches disabled", "evidence": "data/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km-llamacpp-faon-cacheoff-ctx2048-confirm-realistic128-20260704T235714Z.json"}, "manifest": "results/rapid-model-snapshots-b70/nemotron-cascade-2-30b-a3b-q4km/README.md" } ], "views": [ {"id": "nemotron-cascade-2-rapid-tp", "title": "Strict rapid operating point", "subtitle": "Q4_K_M · one B70 · f16 KV · ctx=2048 configured; this one point is not a context curve", "x_label": "tensor parallel cards", "discrete": true, "missing_x": [2, 4], "metrics": ["decode_tok_s", "ttft_ms"], "series": [{"label": "Q4_K_M", "measurement_ids": ["nemotron-cascade-2-q4km-tp1-rapid"], "x_from": "config.tp"}]} ], "coverage_views": [ { "id": "nemotron-cascade-2-quant-by-tp", "label": "quant × TP", "fixed": "Exact measured revision and rapid runtime; TP2/4 are unmeasured gaps.", "row_axis": {"key": "variant", "label": "Quantization", "prefix": ""}, "column_axis": {"key": "tp", "label": "TP", "prefix": "TP"}, "fixed_selectors": {"revision": "nemotron-cascade-2-30b-a3b-931b595", "runtime": "llama.cpp SYCL fdb1db877", "mtp": 0}, "rows": ["Q4_K_M"], "columns": [1, 2, 4], "cells": { "Q4_K_M:1": {"state": "lab-measured", "label": "D50.904 · T449.159", "evidence_id": "nemotron-cascade-2-q4km-tp1-rapid", "packet_id": "rapid-nemotron-cascade-2-30b-a3b-q4km"}, "Q4_K_M:2": {"state": "missing", "label": "no stored lane"}, "Q4_K_M:4": {"state": "missing", "label": "no stored lane"} } } ], "family_closures": [] }