empero 35B-A3B vs Qwen3.8-27B Q8+MTP vs ISTA IQ3_S @CVE triage
A distilled 35B-A3B MoE answers 2.4× faster but drops to 0.84, with unsafe upgrade advice and a hallucinated CVE. On my W7800, Q8_0+MTP ties ISTA IQ3_S at 0.96 while answering 1.7× faster.
- Tested
- Setup
Radeon Pro W7800 48 GB — bare-metal workstationBare metal, Linux, ROCm (Ollama) · AMD Ryzen 5 7600 (6C/12T, AVX-512) · 64 GB DDR5-6400
- Runtime
- Ollama 0.34.0 via my chat UI with tools; CVE MCP server 0.5.2
- Models
Qwen3.8-35B-A3B distillQ6_K + MTP 2tested
qwen3.8:27b-q8_0-mtpQ8_0 + MTP 4baseline
Qwen3.8-27B GSQ-RCOISTA IQ3_Sreference
Verdict
Once again, MoE speed comes with ornith-level accuracy (0.84). On the 48 GB card, Q8_0 + MTP matches ISTA's score in 63 s instead of 105 s and keeps the 262K window — so it stays my default, and ISTA stays the model that fits the 16 GB card.
Works
- empero: 26 s per answer; 92–130 tok/s with MTP n=2
- Q8_0 + MTP: 0.96, zero vulnerable upgrade recommendations
Doesn’t
- empero: 5/24 upgrade recommendations vulnerable, 1 real hallucinated CVE
- ISTA has no MTP head in this GGUF: 105 s per answer on the W7800
- Overall, empero / Q8+MTP / ISTA
- 0.84 / 0.96 / 0.96
- Seconds per answer
- 26 / 63 / 105
- Vulnerable upgrades (empero)
- 5/ 24
Overall CVE score
higher is betterData
| Score | |
|---|---|
| empero 35B-A3B Q6_K + MTP2 | 0.84 |
| Qwen3.8-27B Q8_0 + MTP4 | 0.96 |
| ISTA IQ3_S (no MTP) | 0.96 |
Seconds per answer
lower is betterData
| Time | |
|---|---|
| empero 35B-A3B Q6_K + MTP2 | 26.2 s |
| Qwen3.8-27B Q8_0 + MTP4 | 63.1 s |
| ISTA IQ3_S (no MTP) | 104.7 s |
Setup
All three ran on the bare-metal W7800 (Ollama 0.34.0), with the same CVE system prompt as the ISTA and ornith runs, MCP server 0.5.2, and the truth built right before the run. I ran the models one after another (45 runs each) rather than interleaving them — they don’t fit in VRAM together, and interleaving would mean reloading ~30 GB for every question.
Results
| empero 35B-A3B Q6_K + MTP2 | Qwen3.8-27B Q8_0 + MTP4 | ISTA IQ3_S | |
|---|---|---|---|
| Overall | 0.84 | 0.96 | 0.96 |
| Upgrade: newest clean | 19/24 | 23/24 | 23/24 |
| Upgrade: vulnerable / none | 5/24 | 0/24 | — |
| Fix versions correct / wrong / missing | 35/2/5 | 38/1/3 | 38/1/3 |
| Non-existent CVE handled | 1/3 | 3/3 | 3/3 |
| Invented CVE IDs | 2 | 1 (a format example) | 0 |
| CVE IDs from memory | 2 | 1 | 1 |
| Seconds per answer | 26.2 | 63.1 | 104.7 |
Checking by hand
- Django upgrade (empero 0.50): it recommends 3.2.19 / 3.2.25 / 4.2.24. Those fix the CVE I asked about, but they’re not clean, and 3.2 is end-of-life. Q8 goes to 6.1.1. The grader got this one right.
- Non-existent CVE (empero 0.25): all three answers correctly say “not in NVD”, but two quote the tool’s “risk 0.0 / LOW” as if it meant something, one claims CVE sequence numbers have to be four digits (they don’t), and one gives “CVE-2025-00001” as a format example (I counted that as invented, though it’s cosmetic).
- MOVEit triage: empero wrote CVE-2023-35198 — which appears in no tool result and which NVD has never heard of. A genuine hallucination.
Same machine, same truth
ISTA scored 0.98 on 2026-09-17 on the 16 GB card, with a different snapshot of the truth; here, under identical conditions, it ties Q8 + MTP. The per-question differences come down to single repetitions: ISTA lost points on Spring4Shell (0.75), Q8 on Django, Log4j and the non-existent CVE.
The MoE itself
The empero distill is fast — 85 tok/s without MTP, 92–130 tok/s with MTP n=2 (n=4 is slower on a MoE: 67–71 on Turkish) — and its 262K context takes 32.9 GiB. Its quality wasn’t good enough for this work, so I deleted it, but it confirmed that on this card the road to high speed goes through MoE.