Laguna XS 2.1 vs Qwen 3.6: AMD R9700 AI Pro Benchmark Results

Published: August 4, 2026 | Reading time: 8 minutes

We ran the same prompt through two different models on the same GPU. One kept digging. One stopped at the surface. The numbers don't tell the whole story — but they do start it.

The Setup

GPU	AMD Radeon R9700 AI Pro (32GB VRAM, ROCm 7.2.4)
Runtime	llama.cpp (latest main, AMD FP4 backend)
Context	65,536 tokens (single slot)
Task	Website evaluation: storyreserve.com
Framework	Hermes Agent (same prompt, same toolset)

Raw Performance Numbers

Metric	Laguna XS 2.1	Qwen 3.6 27B
Generation speed	46 tok/s	38 tok/s
VRAM used	~27 GB	~19 GB
Model size (Q4_K_M)	~7.5 GB	~16 GB
Context window	65K	65K
Load time (cold)	~8s	~12s

Laguna is faster. Uses more VRAM for KV cache. Lighter model file. Quicker to warm up.

But Speed Isn't the Story

Here's what happened when both models were asked to evaluate storyreserve.com:

Laguna XS 2.1: The Fast Surface Scan

Homepage fetched in 263ms
Correctly identified as "AI-generated Christian literature"
Stated "legitimate site, no scam indicators"
Stopped there. No endpoint discovery. No work count.

Total steps: 3. Total time: ~12 seconds. Clean. Fast. Incomplete.

Qwen 3.6: The Deep Audit

Homepage fetched
Discovered /data/works.json (1.7MB catalog file)
Counted 114 works across 9 genres
Found production bug: only 1 of 114 works has a buy link
Verified audio endpoint (566MB content range)
Confirmed POST /api/submit is alive

Total steps: 8+. Total time: ~25 seconds. Slower. Thorough. Actionable.

The Behavioral Difference

Same framework. Same tools. Same GPU. Different model behavior:

Behavior	Laguna	Qwen
First action	Fetch homepage, summarize	Fetch homepage, probe structure
When tool returns data	Surface-level interpretation	Hunts for additional endpoints
On 404 error	Loops same failing call	Pivots to sitemap/alternative
Depth instinct	"Good enough"	"What else is there?"

Laguna feels like a planner — efficient, structured, stops when the plan is complete. Qwen feels like a pragmatist — probing, adapting, keeps going until the job is actually done.

When to Use Which

Use case	Recommended	Why
Quick summary, first impression	Laguna	Fast, coherent, gets the gist
Website audit, security review	Qwen	Discovers hidden endpoints, counts real assets
Code review, multi-file analysis	Qwen	Cross-references, follows imports, finds edge cases
Chatbot, customer-facing Q&A	Laguna	Fast response, lower latency, friendlier tone
Data extraction, catalog processing	Qwen	Handles large JSON, structured counting, validation

The ROCm Angle

Both models run on AMD's ROCm stack — not CUDA. The R9700 AI Pro is a 32GB card with dedicated AI acceleration, and llama.cpp's AMD FP4 backend gives us competitive performance without the NVIDIA tax.

Key ROCm-specific notes:

ROCm 7.2.4 supports Flash Attention on gfx1201 (R9700's architecture)
KV cache quant to q8_0 saves ~30% VRAM without measurable quality loss
Single-slot mode (1 parallel user) maximizes context depth over throughput

We don't need dual GPUs. One 32GB Radeon handles 65K context with room to spare — something two RTX 3060s (12GB each) can't match without splitting layers across PCIe.

Bottom Line for Businesses

If you're choosing local AI models for your business:

Speed shoppers: Laguna wins on tok/s and latency. Good for customer-facing chat.
Depth shoppers: Qwen wins on audit quality and recovery. Good for due diligence and automation.
Both: You can run both. Same GPU. Different ports. Pick the brain for the job.

We run Laguna for quick tasks and Qwen for audits. Same box. No cloud. Your data never leaves the server.

About BizFlowAI: We build custom AI automation for small businesses powered by local AI on AMD and NVIDIA hardware. Talk to us about your stack.