Released as part of the NOESIS Professional Multilingual Dubbing Automation Platform (framework: DHCF-FNO — Deterministic Hybrid Control Framework for Frozen Neural Operators).

Founder: Ilia Bolotnikov
Organization: AMAImedia.com
X (Twitter): @AMAImediacom
LinkedIn: Ilia Bolotnikov
Telegram: @djbionicl
NOESIS version: v16.1
Release date: 2026-08

NOESIS-Qwopus3.5-4B-v3-Supervisor-LongCtx-NF4

Role: Long-context Supervisor — 6GB-runtime NF4 build (multi-segment QC, orchestration, cross-stage review, batch best-of-N) for the dubbing pipeline. Quantized from NOESIS-Qwopus3.5-4B-v3-Supervisor-LongCtx-BF16 (bnb load_in_4bit, nf4, double_quant, compute=bf16). PRIMARY artifact = the BF16 sibling; GGUF Q8_0 = laptop. Equivalent to base-NF4 + LoRA adapter nt346_longctx_qwopus4b, pre-merged.

Test results (2026-06-17)

Score measured on the matching Q8_0 GGUF (NF4 is the same merged weights at 6GB runtime):

Supervisor eval — eval_longctx_supervisor_v1.py (12 tests, grammar-constrained)

• 11/12 (Q8_0). Only failure = L1_mixed (fail-safe over-escalation). Prior best long-ctx (LFM2.5-7.5B) = 7/12.

Translation — FLORES devtest (n=20, no-think)

• eng→rus chrF++ 51.5 / BLEU 25.6 ; eng→cmn 32.0 / 7.4 ; AVG 41.7 / 16.5.

Comparison

Model	Supervisor-12	Translate AVG chrF++/BLEU
This (4B-LongCtx)	11/12	41.7 / 16.5
base 4B (no LoRA)	5/12	42.4 / 15.4
Qwopus3.5-9B-Translate Q4	5/12	43.8 / 16.5

Speed (RTX 3060 Laptop 6GB, GPU, Q8_0 sibling, 33/33 layers)

• gen 53.5 tok/s, prompt eval 366 tok/s (vs 9B-Translate Q4: 49.1 / 307).

All-rounder: full supervision + translation parity with the base, smaller footprint. Written: 2026-06-17