Petronella Technology Group · Fleet LLM Benchmark
Independent · measured on-fleet · updated July 15, 2026

The private-AI model benchmark for work that can't leave your building.

We evaluate open-weight and cloud LLMs the way a compliance-bound firm actually uses them — coding, hard debugging, cited research, agentic tool-use and long-form writing — judged by a three-model frontier panel and run on Petronella's own GPU fleet, so CUI and privileged data never touch a third party.

70+
models evaluated
3-judge
frontier panel (median)
H200
+ Blackwell & GB10 fleet
67.3%
SWE-bench Verified, local
Latest study · July 15, 2026

Gemma-4-26B-A4B vs. the Qwen3.6 family — difficulty decides the winner.

Under a three-model frontier panel (GPT-5.4 + Claude Opus-4.8 + GLM-5.2, median score on identical generations, N=5), the answer splits by task difficulty. Qwen wins routine coding; Gemma wins the hard coding set decisively and is the stronger agent — verified end-to-end through our live agent gateway.

Gemma-4-26B-A4B (NVFP4)
Google · 26B MoE, ~4B active
  • Hard coding 0.915 — wins
  • Best tool-caller (55 correct calls)
  • Multi-turn agentic 1.00 pass@max
vs
Qwen3.6 family
Alibaba · 35B-A3B / 27B dense
  • Easy coding 0.98 — wins
  • Hard coding 0.78–0.81
  • Agentic pass@max 0.778

Serving notes that mattered: Gemma-4 needs vLLM --tool-call-parser gemma4 (the wrong parser makes zero tool calls) and --max-model-len 131072 to fit the agent gateway's ~58k-token system prompt. NVFP4 is near-lossless for Gemma-4 but not for Qwen3.6-35B (a long-form runaway appears — use FP8). Full findings →

Recommendations

Best model for your hardware

The single most decision-useful view: the strongest model we found for each fleet configuration and role, with the trade-offs that decided it.

4-way H200 NVL575 GB pool
Ornith-1.0-397B-FP8 LOCAL MIT
Leads the hard coding set outright (0.971, 20/20 every rep) and scored a perfect 1.000 blog across three topics; research 0.91 at 120 t/s. Alternatives: MiniMax-M3-MXFP8, GLM-5.2-NVFP4 (research leader 0.94).
2-way H200 NVL287 GB pool
Qwen3-Coder-Next-FP8 LOCAL
80B-A3B, Apache-2.0. Best coder in class on the hard set (0.939 at 2.6 s/case), 85.2% on the 88-case tool corpus, 178 t/s single / 2,789 t/s at c=32. Long-form alt: MiniMax-M2.7-NVFP4 (blog 0.94).
Single 96 GB cardRTX PRO 6000 Blackwell
Qwen3.6-35B-A3B-FP8 LOCAL
Best all-rounder that fits one card: coding 0.975, wins research and cited-RAG. No displacer found. For the hard set specifically, Gemma-4-26B-A4B (0.915) is the more robust choice.
Long context524K tokens
DeepSeek-V4-Flash LOCAL
Coding 0.964, blog 0.944, and 67.3% on SWE-bench Verified measured locally — essentially tying cloud DeepSeek-V4-pro. Also the fleet's validated free grader.
Voice tool-callerlatency-critical
gpt-oss-20b LOCAL
125/125 correct declines and 0% tool hallucination on the 88-case decline-safety corpus. The metric that matters for a voice agent is not calling the wrong tool — and it never does.
Air-gapped edgeappliance / laptop
Ministral-3-8B · Granite-4.1-8B · Gemma-4-e4b LOCAL
Compact models that still drive tools on-device, for ComplianceArmor appliances where the alternative is no local model at all. Mistral models need --tool-call-parser mistral.
Category leaders

Who tops each task

Best score per model, cross-judged (generator never grades itself). Cloud models are shown as reference lines; everything marked LOCAL runs entirely on the PTG fleet.

Coding — core

22 PTG ops/SEO/debug tasks · temp 0
1Claude Opus-4.7 CLOUD0.991
2Qwen3.6-27B dense (BF16) LOCAL0.990
3gpt-oss-20b LOCAL0.989
4DeepSeek-V4-Flash-DSpark LOCAL0.986

Coding — hard set

20 harder spec/debug/multi-file cases
1Ornith-1.0-397B-FP8 LOCAL0.971
2GLM-5.2 FP8 (reference) CLOUD0.952
3GLM-5.2 NVFP4 LOCAL0.942
4Qwen3-Coder-Next-80B LOCAL0.939

Research & reasoning

27 cases: summarization, multi-hop, citations
1Gemma-4-31B / 26B-A4B LOCAL0.963
2Qwen3-Coder-Next-80B LOCAL0.963
3Nex-N2-Pro (NVFP4) LOCAL0.963
4GLM-5.2 LOCAL0.94

Blog writing

3,000-word SEO posts · 10-criteria + judge
1Qwen3.6-27B dense LOCAL1.000
2Nex-N2-Pro / Ornith-397B LOCAL1.000
3DeepSeek-V4-Flash LOCAL0.944
4Gemma-4-26B-A4B LOCAL0.789

Full tables with per-rep ranges, latency, temperature and dates →

Recent findings

What we've learned

The headline results driving PTG's private-AI deployment decisions — every number measured on our own hardware with a real evaluation harness, not vendor marketing.

Throughput · July 2026

For single-user work, a laptop iGPU now trades blows with a Grace-Blackwell box.

Gemma-4-26B-A4B (4B active) on an AMD Strix Halo laptop — Radeon 8060S iGPU, unified LPDDR5X — decodes 63 t/s single-stream on llama.cpp Vulkan (43 t/s via Ollama), clearly ahead of the NVIDIA GB10's best vLLM single-stream (44.4 t/s). It is model-dependent — Qwen3.6's MTP path keeps the GB10 ahead there (95 vs 65 t/s) — and the same 96 GB unified pool loads gpt-oss-120b whole: a 120B-class MoE at 37 t/s on a laptop. The GB10 keeps the serving crown: vLLM batching scales it to 554 t/s at 32 concurrent streams, where llama.cpp cannot follow.

SWE-bench Verified · June 2026

A near-cloud, CUI-capable coder already runs on our fleet.

On the official SWE-bench Verified harness (stratified-150, no LM judge), DeepSeek-V4-Flash local scored 67.3% — essentially tying cloud DeepSeek-V4-pro (69.6%), beating the fleet's Qwen3.6 (55.3%), about 15 points under Claude Opus-4.8 (82.0%). We own a private coder within reach of the frontier.

July 2026

NVFP4 4-bit is lossless — for some models, not all.

Unsloth's NVFP4 quants hold full quality for Gemma-4 (coding and research within noise of BF16) at half the VRAM on Blackwell. But Qwen3.6-35B develops a long-form runaway at NVFP4 that FP8 doesn't — so quant choice is per-model, not universal.

July 2026

Serving config is a first-class correctness variable.

Gemma-4 with the wrong vLLM tool parser emitted zero tool calls — it looked agentically incapable when it was a one-line misconfiguration. With --tool-call-parser gemma4 it became the best tool-caller in the set. Both models then passed real end-to-end runs through our live agent gateway.

Methodology · May 2026

Never let a model grade its own family.

Self-judging inflates scores 20–37%. Our cross-judge gate forbids same-family grading, and headline studies now use a three-model frontier panel (median of GPT-5.4, Opus-4.8, GLM-5.2) — which repeatedly caught a single judge under-rating one contestant by ~0.3.

Methodology

How the scores are produced

Task-appropriate, cross-judged. Each model runs a fixed corpus per capability — 22 core + 20 hard coding tasks, 27 research cases, an 88-case tool decline-safety corpus, agentic multi-turn tasks, and a 3,000-word blog with a 10-criteria rubric. Coding and research combine deterministic auto-checks (does the code run, are the citations right) with an LM judge; blog combines structural gates with a judge. The generator is never the judge, and never shares its model family with the judge.

Frontier judge panel. Headline comparisons are graded by three independent frontier models — GPT-5.4 (OpenAI), Claude Opus-4.8 (Anthropic) and GLM-5.2 (Zhipu) — scoring identical saved generations, with the median taken per case so no single judge's bias can decide a result. DeepSeek-V4-Flash serves as a validated free grader for bulk work.

Measured on-fleet. Everything is timed and scored on Petronella's own GPUs — the point is that a private, compliance-bound firm can run capable AI on privileged and CUI data without sending a single token to a third party.

6× H200 NVL 143 GBRTX PRO 6000 Blackwell 96 GB NVIDIA GB10 Grace-BlackwellAMD Strix Halo (ROCm) Apple M1 Max / M5 (MLX)vLLM · NVFP4 / FP8 / int4