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-BF16

Role: Long-context Supervisor — PRODUCTION (multi-segment QC, orchestration plan, cross-stage review, batch best-of-N) for the dubbing pipeline. From: base NOESIS-Qwopus3.5-4B-v3 + plain-LoRA nt346_longctx_qwopus4b (CCE, max_len 768, completion-masked) merged. Trained on NF4 base, merged into BF16. BF16 = PRIMARY. Siblings: -NF4 (6GB runtime), -Q8_0.gguf (laptop), LoRA adapter.

Test results (2026-06-17)

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

Quant	Size	Score
Q8_0	4.17 GB	11/12
IQ2_XXS	1.43 GB	8/12 (2-bit cliff: under-escalates on 248K vocab — not recommended)

Only failure at 11/12 = L1_mixed (aggregates trunc+speed mix to reject vs retry; fail-safe). Prior best long-ctx model (LFM2.5-7.5B) was 7/12.

Translation — FLORES devtest (chrF++/BLEU, n=20, no-think)

Direction	chrF++	BLEU
eng→rus	51.5	25.6
eng→cmn	32.0	7.4
AVG	41.7	16.5

Comparison vs dedicated models (same FLORES n=20)

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

Takeaway: gains full supervision (+6 over base) with no loss of translation (parity with base; within ~2 chrF of the dedicated 9B translator at < half the params). Strong all-rounder.

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

• Q8_0: gen 53.5 tok/s, prompt eval 366 tok/s. (vs 9B-Translate Q4: 49.1 / 307.)

Eval on GPU via standalone llama-completion.exe -ngl 99 (pip llama_cpp is CPU-only). Written: 2026-06-17