# Complete Multi-Axis Analysis — Agent Model Selection on Mac Mini M4 16GB **Date:** 2026-04-08 | **Tests:** 6 agent tasks × 15+ configurations | **Vision:** Falcon Perception v2 --- ## Axis Overview ``` 5 axes tested across 15+ model configurations: Axis 1: Model Family ─── Qwen3.5-9B vs Gemma4-E4B vs Qwen3VL-8B vs Bonsai vs others Axis 2: Censoring ────── Base (censored) vs Uncensored (Balanced vs Aggressive) Axis 3: Quantization ─── Q4 vs Q5 vs Q6 vs Q8 vs 1-bit vs BF16/4bit-MLX Axis 4: Backend ──────── llama.cpp (GGUF) vs mlx_vlm (MLX) vs Ollama vs PrismML Axis 5: Vision ──────── with mmproj vs without vs Falcon Perception detector ``` --- ## AXIS 1: Model Family ``` Score Speed Multi-step Form Fill Vision Min Params (/6) (tok/s) Chains? Works? Detect? for Agent ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ Qwen3.5-9B (9B dense) 5.0 13.5 ✅ Yes ✅ Yes ✅ FP 9B Gemma4 E4B (4B MoE) 5.0 24.5 ✅ Yes ✅ Yes ✅ FP 4B active Qwen3VL-8B (8B dense) 3.0 16.2 ⚠️ Crashes ❌ Crashed ✅ FP 8B Bonsai-8B (8B 1-bit) 1.0 48.8 ❌ 1 turn ❌ No ❌ No 8B (degraded) LFM2.5-Nova (1.2B) 0.0 118.4 ❌ 1 turn ❌ No ❌ No 1.2B FunctionGemma (270M) 0.0 197.0 ❌ Loops ❌ No ❌ No 270M Qwopus-27B Q3 (27B) OOM OOM ❌ OOM ❌ OOM ❌ OOM 27B (too big) INSIGHT: The agent capability cliff is at ~4B active params. Below 4B: can format tool calls but cannot chain them. Above 4B: can navigate, search, fill forms, use vision. Model FAMILY matters less than active parameter count. Gemma4 (4B active MoE) matches Qwen3.5 (9B dense). ``` ### Family Comparison (best variant of each) ``` Score vs Speed: 5.0 │ ★ Qwen3.5 Unc Q6K ★ Gemma4 Unc Q5KP │ (13.5 tok/s) (24.5 tok/s) 4.0 │ │ 3.0 │ ★ Qwen3VL Bal Q6K │ (16.2 tok/s) 2.0 │ │ 1.0 │ ★ Bonsai │ (48.8 tok/s) 0.0 │ ★ LFM2.5 ★ FuncGemma └──────┬──────────┬──────────┬──────────┬──────────┬──────────┬ 10 20 30 50 118 197 Speed (tok/s) ``` --- ## AXIS 2: Censoring (Base vs Uncensored) ``` ┌────────────────────────┬───────────┬───────────┬───────────┐ │ │ Base │Uncensored │Uncensored │ │ │ (censor) │ Balanced │ Aggressive│ ├────────────────────────┼───────────┼───────────┼───────────┤ │ Qwen3.5-9B Q4 score │ 4.5/6 │ — │ 3.5/6 ↓ │ │ Qwen3.5-9B Q6 score │ — │ — │ 5.0/6 ★ │ │ Gemma4 E4B Q6 score │ — │ — │ 4.5/6 │ │ Gemma4 E4B Q5 score │ — │ — │ 5.0/6 ★ │ │ Qwen3VL-8B Q6 score │ — │ 3.0/6 │ — │ │ │ │ │ │ │ Tool call reliability │ Good │ Similar │ Better │ │ Refusal rate │ Some │ Zero │ Zero │ │ Stop discipline │ Good │ Similar │ Better │ │ Form fill success │ ✅ │ — │ ✅ │ └────────────────────────┴───────────┴───────────┴───────────┘ INSIGHT: Uncensoring does NOT directly improve agent tasks. The improvement we see (3.5→5.0 for Qwen Q4→Q6) is from QUANTIZATION, not censoring. Base Qwen3.5-9B Q4_K_XL (4.5/6) vs Uncensored Q4_K_M (3.5/6) → uncensored is WORSE at Q4. The uncensored Q6_K (5.0/6) wins because Q6 > Q4, not because uncensored > censored. Uncensoring helps with: content generation, creative tasks. Uncensoring neutral for: tool calling, agent planning. Uncensoring harmful at: Q4 level (less reliable stops). ``` --- ## AXIS 3: Quantization ### Gemma4 E4B Uncensored — Quant Ladder ``` ┌─────────┬────────┬────────┬───────┬───────┬───────┬───────┐ │ Quant │ Size │ Memory │Speed │Score │Stops │Tools │ │ │ (GB) │ (GB) │tok/s │ (/6) │ (/6) │ total │ ├─────────┼────────┼────────┼───────┼───────┼───────┼───────┤ │ Q5_K_P │ 5.4 │ 6.3 │ 24.5 │ 5.0 ★ │ 3 ★ │ 50 │ │ Q6_K_P │ 5.8 │ 6.7 │ 23.1 │ 4.5 │ 2 │ 52 ★ │ │ Q8_K_P │ 7.6 │ 8.5 │ 19.0 │ 4.0 │ 2 │ 41 │ │ 4bit MLX│ ~5.0 │ 5.4 │ 35.0★ │ 4.5 │ 1 │ 29 │ └─────────┴────────┴────────┴───────┴───────┴───────┴───────┘ Speed vs Score: Score 5.0 │ ★ Q5_K_P │ 4.5 │ ★ Q6_K_P ★ 4bit-MLX │ 4.0 │ ★ Q8_K_P │ 3.5 │ └──┬────────┬────────┬────────┬──── 19 23 25 35 Speed (tok/s) INSIGHT: Higher quantization ≠ better agent performance! Q5 > Q6 > Q8 for agent tasks. Why? 1. Q5 is FASTER → more turns per timeout → more work done 2. Q5 has enough precision for tool call JSON formatting 3. Q8 wastes memory on precision the model can't use for reasoning 4. The bottleneck is REASONING DEPTH (4B params), not PRECISION SWEET SPOT: Q5_K_P — best score, fastest, smallest ``` ### Qwen3.5-9B Uncensored — Quant Ladder ``` ┌─────────┬────────┬────────┬───────┬───────┬───────┐ │ Quant │ Size │ Memory │Speed │Score │Stops │ ├─────────┼────────┼────────┼───────┼───────┼───────┤ │ Q4_K_M │ 5.2 │ 6.1 │ 16.7 │ 3.5 │ 0 │ │ Q6_K │ 6.9 │ 7.8 │ 13.5 │ 5.0 ★ │ 2 ★ │ │ Q8_0 │ 8.9 │ ~10 │ ~10 │ N/A │ N/A │ │ Base Q4 │ 5.6 │ 6.5 │ 10.0 │ 4.5 │ 2 │ └─────────┴────────┴────────┴───────┴───────┴───────┘ INSIGHT: For Qwen (9B dense), HIGHER quant IS better. Q4 (3.5) → Q6 (5.0) = massive improvement. The 9B dense model has more capacity that benefits from higher precision. Unlike the 4B MoE Gemma4 where the bottleneck is active params not precision. SWEET SPOT: Q6_K — best score, fits 16GB ``` ### Cross-Model Quant Pattern ``` FINDING: Optimal quantization depends on active parameter count. Dense 9B (Qwen): Q4 < Q6 ★ (precision limited, higher=better) MoE 4B (Gemma4): Q5 ★ > Q6 > Q8 (speed limited, faster=better) 1-bit 8B (Bonsai): Degraded regardless (1-bit destroys reasoning) Rule of thumb: • Dense models: use highest quant that fits • MoE models: use mid quant (Q5) for best speed/quality • Ultra-small quant ( Gemma4 MLX (35 tok/s) because reliability > raw speed for agent tasks. ``` --- ## AXIS 5: Vision Configuration ``` ┌─────────────────────────┬──────────┬──────────────────────────────────────┐ │ Configuration │ Score │ How vision works │ ├─────────────────────────┼──────────┼──────────────────────────────────────┤ │ mmproj only (VLM sees │ 4.5/6 │ LLM sees screenshots directly. │ │ screenshots directly) │ │ Slow: 30-40s to process each image. │ │ │ │ Inaccurate: misidentifies objects. │ │ │ │ Uses all context for image tokens. │ ├─────────────────────────┼──────────┼──────────────────────────────────────┤ │ Falcon Perception only │ 4.5/6 │ Dedicated 0.6B detector. │ │ (no mmproj, text-only │ │ Fast: 2s per detection. │ │ LLM + vision_detect) │ │ Returns pixel coordinates. │ │ │ │ LLM never sees images — saves ctx. │ ├─────────────────────────┼──────────┼──────────────────────────────────────┤ │ mmproj + Falcon (both) │ 5.0/6 ★ │ Best of both worlds: │ │ │ │ • LLM sees screenshots (understands │ │ │ │ page layout, text, context) │ │ │ │ • Falcon detects objects precisely │ │ │ │ (captcha grids, UI elements) │ │ │ │ • 3-layer adaptive pipeline routes │ │ │ │ to best method per task │ ├─────────────────────────┼──────────┼──────────────────────────────────────┤ │ No vision at all │ 1-2/6 │ Model can only navigate + extract │ │ (text-only, no images) │ │ text. Can't interact with visual UI. │ └─────────────────────────┴──────────┴──────────────────────────────────────┘ Falcon Perception 3-Layer Pipeline: Layer 1 (Router): "captcha" → detection (2s) "measure" → segmentation (5-9s) Layer 2 (Present): default → coordinates text (fastest) verify → cropped images for VLM spatial → Set-of-Marks overlay Layer 3 (Extract): coords: "[1] center=(211,155)" crops: 3 JPEG images per detection overlay: annotated full image INSIGHT: mmproj + Falcon together > either alone. mmproj lets the LLM understand page context. Falcon gives pixel-accurate object detection. The combination scores 5.0/6 vs 4.5/6 for either alone. Memory cost: Falcon = 1.5GB always loaded. Speed cost: 2s per vision_detect call (negligible vs LLM). ``` --- ## THE COMPLETE PICTURE ### All Tested Configurations Ranked ``` Rank Model + Config Score Speed Memory Proxy Backend ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1. Gemma4 Unc Q5_K_P +mm +FP 5.0/6 24.5 6.3G No llama.cpp 2. Qwen3.5 Unc Q6_K +mm +FP 5.0/6 13.5 7.8G No llama.cpp 3. Gemma4 Unc Q6_K_P +mm +FP 4.5/6 23.1 6.7G No llama.cpp 4. Gemma4 Base 4bit +FP 4.5/6 35.0 5.4G YES mlx_vlm 5. Qwen3.5 Base Q4_K_XL +mm +FP 4.5/6 10.0 6.5G No llama.cpp 6. Gemma4 Unc Q8_K_P +mm +FP 4.0/6 19.0 8.5G No llama.cpp 7. Qwen3.5 Unc Q4_K_M +mm +FP 3.5/6 16.7 6.1G No llama.cpp 8. Qwen3VL Bal Q6_K +mm +FP 3.0/6 16.2 7.4G No llama.cpp 9. Bonsai-8B 1-bit +FP 1.0/6 48.8 1.5G No PrismML 10. LFM2.5-Nova 1.2B Q4 0.0/6 118.4 0.8G No llama.cpp 11. FunctionGemma 270M Q8 0.0/6 197.0 0.3G No llama.cpp 12. Qwopus-27B Q3_K_S OOM OOM 14.0G No llama.cpp +mm = with mmproj | +FP = with Falcon Perception ``` ### Key Decisions ``` ┌─────────────────────────────────────────────────────────────────────┐ │ DECISION FRAMEWORK │ │ │ │ Q: Which model? │ │ A: Gemma4 E4B Uncensored Q5_K_P │ │ • Tied #1 at 5.0/6, fastest winner (24.5 tok/s) │ │ • 1.5GB smaller than Qwen alternative │ │ • No proxy needed (GGUF on llama.cpp) │ │ • Form filling works (the Gemma4 MLX version couldn't) │ │ │ │ Q: Which quantization? │ │ A: For MoE (Gemma4): Q5_K_P (speed > precision) │ │ For Dense (Qwen): Q6_K (precision > speed) │ │ Never Q8+ (wastes memory, slower, no quality gain) │ │ Never Q4 or below (breaks multi-step reasoning) │ │ │ │ Q: Which backend? │ │ A: llama.cpp (stock homebrew) for everything │ │ MLX only if you need 35 tok/s AND accept proxy complexity │ │ │ │ Q: Censored or uncensored? │ │ A: Doesn't matter for agent tasks. Quality comes from quant. │ │ Uncensored helps if tasks involve restricted content. │ │ │ │ Q: What about vision? │ │ A: Always use mmproj + Falcon Perception together (5.0/6) │ │ Either alone = 4.5/6. Both = synergy. │ │ │ │ Q: Can we get 6/6? │ │ A: T2 (DDG) and T6 (CAPTCHA) timeout — model works but │ │ doesn't stop in time. Improving stop_loop instructions │ │ or extending timeouts would push to 5.5-6.0/6. │ │ The model DOES the work — it just keeps going. │ └─────────────────────────────────────────────────────────────────────┘ ``` ### Counter-Intuitive Findings ``` 1. BFCL ≠ Agent Capability Bonsai: 73% BFCL but 1.0/6 agents (single-turn ≠ multi-turn) 2. Bigger Quant ≠ Better Agent (for MoE) Gemma4: Q5 (5.0) > Q6 (4.5) > Q8 (4.0) — speed wins over precision 3. Bigger Quant = Better Agent (for Dense) Qwen: Q4 (3.5) < Q6 (5.0) — precision matters for 9B dense 4. Uncensored ≠ Better Agent Same model, same quant: uncensored doesn't improve tool calling 5. Faster Backend ≠ Better Results Gemma4 MLX (35 tok/s, 4.5) < Gemma4 GGUF (24 tok/s, 5.0) Proxy complexity and message format issues hurt more than speed helps 6. 270M → 9B Speed Inversion FunctionGemma (197 tok/s, 0/6) < Gemma4 (24 tok/s, 5/6) 10x faster but completely useless for the actual task 7. 4B Active Params = 9B Dense Gemma4 MoE (4B active) matches Qwen (9B dense) on agent tasks MoE architecture is incredibly efficient for tool calling ``` --- ## Memory & Disk Reference ``` 16GB Mac Mini M4 — Memory Budget: Gemma4 Q5_K_P (winner): Model: 5.4 GB ████████████████████████████ mmproj: 0.9 GB █████ Falcon: 1.5 GB ████████ GUA+Brwsr: 0.8 GB ████ OS: 3.0 GB ████████████████ KV Cache: ~1.5 GB ████████ ───────────────────────────────── Total: 13.1 GB Free: 2.9 GB ✅ Qwen Q6_K (runner-up): Model: 6.9 GB ████████████████████████████████████ mmproj: 0.9 GB █████ Falcon: 1.5 GB ████████ GUA+Brwsr: 0.8 GB ████ OS: 3.0 GB ████████████████ KV Cache: ~1.5 GB ████████ ───────────────────────────────── Total: 14.6 GB Free: 1.4 GB ⚠️ tight ``` --- *15+ configurations tested across 5 axes, 90+ individual test runs.* *Mac Mini M4 16GB (Dyson), April 3-8 2026.*