diff --git a/Makefile b/Makefile
index 13ac948..629119a 100644
--- a/Makefile
+++ b/Makefile
@@ -48,6 +48,9 @@ optimize-kernel: ## Configure kernel boot parameters
 optimize-tuned: ## Switch to accelerator-performance profile
 	@bash scripts/optimize/tuned-profile.sh
 
+optimize-power: ## Apply Phase 2 tuning (ryzenadj, sysctl, THP, RADV)
+	@bash scripts/optimize/power-profile.sh
+
 optimize-vram: ## BIOS VRAM guidance + GTT verification
 	@bash scripts/optimize/vram-gtt.sh
 
diff --git a/bin/optimize b/bin/optimize
index 406400b..3443e36 100755
--- a/bin/optimize
+++ b/bin/optimize
@@ -6,24 +6,31 @@ SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 
 case "${1:---all}" in
     --all|-a)
-        echo "Running optimization walkthrough..."
+        echo "Running full optimization walkthrough..."
+        echo ""
+        echo "=== Phase 1: Core System ==="
         bash "$SCRIPT_DIR/scripts/optimize/tuned-profile.sh"
         bash "$SCRIPT_DIR/scripts/optimize/kernel-params.sh"
         bash "$SCRIPT_DIR/scripts/optimize/vram-gtt.sh"
         echo ""
+        echo "=== Phase 2: System Tuning ==="
+        bash "$SCRIPT_DIR/scripts/optimize/power-profile.sh"
+        echo ""
         bash "$SCRIPT_DIR/scripts/optimize/verify.sh"
         ;;
     --kernel|-k)  exec bash "$SCRIPT_DIR/scripts/optimize/kernel-params.sh" ;;
     --tuned|-t)   exec bash "$SCRIPT_DIR/scripts/optimize/tuned-profile.sh" ;;
     --vram|-v)    exec bash "$SCRIPT_DIR/scripts/optimize/vram-gtt.sh" ;;
+    --power|-p)   exec bash "$SCRIPT_DIR/scripts/optimize/power-profile.sh" ;;
     --verify)     exec bash "$SCRIPT_DIR/scripts/optimize/verify.sh" ;;
     --rollback)   exec bash "$SCRIPT_DIR/scripts/optimize/rollback.sh" ;;
     *)
-        echo "Usage: optimize [--all|--kernel|--tuned|--vram|--verify|--rollback]"
-        echo "  --all       Full optimization walkthrough (default)"
+        echo "Usage: optimize [--all|--kernel|--tuned|--vram|--power|--verify|--rollback]"
+        echo "  --all       Full optimization walkthrough (Phase 1 + 2)"
         echo "  --kernel    Configure kernel boot parameters"
         echo "  --tuned     Switch tuned profile"
         echo "  --vram      BIOS VRAM + GTT guidance"
+        echo "  --power     Phase 2: ryzenadj, sysctl, THP, RADV"
         echo "  --verify    Post-optimization checklist"
         echo "  --rollback  Revert changes"
         exit 1
diff --git a/configs/models.conf b/configs/models.conf
index 1edfba4..ea0b05e 100644
--- a/configs/models.conf
+++ b/configs/models.conf
@@ -21,3 +21,10 @@ qwen3.5-27b-q4|unsloth/Qwen3.5-27B-GGUF|Qwen3.5-27B-Q4_K_M.gguf|17|dense|Dense 2
 qwen3.5-35b-a3b-q4|unsloth/Qwen3.5-35B-A3B-GGUF|Qwen3.5-35B-A3B-UD-Q4_K_L.gguf|19|moe|MoE 35B, 3B active, Unsloth dynamic
 qwen3.5-35b-a3b-q8|unsloth/Qwen3.5-35B-A3B-GGUF|Qwen3.5-35B-A3B-Q8_0.gguf|37|moe|MoE 35B Q8, near-full precision
 nemotron-30b-a3b-q4|lmstudio-community/NVIDIA-Nemotron-3-Nano-30B-A3B-GGUF|NVIDIA-Nemotron-3-Nano-30B-A3B-Q4_K_M.gguf|23|moe|Nemotron MoE 30B, 3B active
+nemotron-cascade2-q8|bartowski/nvidia_Nemotron-Cascade-2-30B-A3B-GGUF|nvidia_Nemotron-Cascade-2-30B-A3B-Q8_0.gguf|31|moe|Nemotron Cascade 2, Mamba-2 hybrid
+
+# ── Coding models ─────────────────────────────────────────
+qwen3-coder-30b-a3b-q6|unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF|Qwen3-Coder-30B-A3B-Instruct-UD-Q6_K_XL.gguf|26|moe|Agentic coding MoE, pure Transformer
+
+# ── Draft models (speculative decoding) ───────────────────
+qwen3.5-0.8b-q8-draft|unsloth/Qwen3.5-0.8B-GGUF|Qwen3.5-0.8B-Q8_0.gguf|0.8|draft|Draft for Qwen3.5 speculative decoding
diff --git a/configs/ryzenadj-llm.service b/configs/ryzenadj-llm.service
new file mode 100644
index 0000000..1112ab8
--- /dev/null
+++ b/configs/ryzenadj-llm.service
@@ -0,0 +1,14 @@
+[Unit]
+Description=Apply RyzenAdj power limits for LLM inference
+After=multi-user.target
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000 --apu-slow-limit=85000
+RemainAfterExit=yes
+
+# Re-apply after resume from sleep/hibernate (HP firmware resets limits)
+ExecStartPost=/bin/sleep 2
+
+[Install]
+WantedBy=multi-user.target
diff --git a/configs/ryzenadj-resume.service b/configs/ryzenadj-resume.service
new file mode 100644
index 0000000..37d8ad7
--- /dev/null
+++ b/configs/ryzenadj-resume.service
@@ -0,0 +1,10 @@
+[Unit]
+Description=Re-apply RyzenAdj power limits after resume
+After=suspend.target hibernate.target hybrid-sleep.target suspend-then-hibernate.target
+
+[Service]
+Type=oneshot
+ExecStart=/usr/local/bin/ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000 --apu-slow-limit=85000
+
+[Install]
+WantedBy=suspend.target hibernate.target hybrid-sleep.target suspend-then-hibernate.target
diff --git a/docs/optimization-log.md b/docs/optimization-log.md
index 1da4d57..7a33e73 100644
--- a/docs/optimization-log.md
+++ b/docs/optimization-log.md
@@ -37,38 +37,42 @@ Living document tracking what was applied, tested, and the actual results. Each
 
 ## Phase 2: System Tuning
 
-### 2.1 RyzenAdj 85W PPT
+### 2.1 RyzenAdj PPT Increase
 
-- **Date**: PENDING
-- **Change**: `sudo ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000`
-- **Expected**: +12-19% CPU/GPU throughput (community data from Strix Halo Wiki)
-- **Benchmark**: Not yet run
-- **Notes**: HP ZBook ships at 60W. 85W is the community-recommended sweet spot.
-- **Verdict**: PENDING
+- **Date**: 2026-03-30
+- **Change**: `sudo ryzenadj --stapm-limit=85000 --fast-limit=85000 --slow-limit=85000 --apu-slow-limit=85000`
+- **Result**: STAPM raised from 59W→81W. PPT Fast raised to 81W. **However, PPT SLOW and APU SLOW stuck at 70W** — HP ZBook BIOS EC overrides these limits. Effective sustained power: ~70W (was ~59W).
+- **Benchmark**: `data/benchmarks/qwen35-shootout-v2-*` (Vulkan, q4_0 KV, pp2048/tg1024)
+  - UD-Q4_K_L: **57.0 t/s** (was ~39 t/s before RyzenAdj = **+46%**)
+  - UD-Q4_K_XL: **56.4 t/s**
+  - Q8_0: **51.4 t/s** (was ~39-41 t/s before = **+25%**)
+- **Thermals**: 70-73C under load, 30C headroom. Cooling handles it easily.
+- **Notes**: Settings are volatile (reset on reboot/sleep). Use `sudo make optimize-power` or install systemd service for persistence. HP firmware hard-caps slow PPT at 70W regardless.
+- **Verdict**: KEEP — significant real-world improvement despite HP firmware limit
 
 ### 2.2 VM Sysctl Tuning
 
-- **Date**: PENDING
-- **Change**: `vm.swappiness=1, vm.dirty_ratio=40, vm.max_map_count=500000`
-- **Expected**: Prevent model weight eviction, reduce I/O disruption
-- **Benchmark**: Not yet run
-- **Verdict**: PENDING
+- **Date**: 2026-03-30
+- **Change**: `vm.swappiness=1, vm.dirty_ratio=40, vm.dirty_background_ratio=10, vm.max_map_count=500000, vm.zone_reclaim_mode=0`
+- **Applied via**: `sudo make optimize-power` (persists to `/etc/sysctl.d/99-llm-inference.conf`)
+- **Notes**: Hard to isolate impact — applied together with other Phase 2 changes. Prevents model weight eviction and I/O disruption.
+- **Verdict**: KEEP — low risk, persists across reboots
 
 ### 2.3 Transparent Huge Pages
 
-- **Date**: PENDING
-- **Change**: `transparent_hugepage=always`
-- **Expected**: Faster model load time, possible 1-5% tg improvement from reduced TLB misses
-- **Benchmark**: Not yet run
-- **Verdict**: PENDING
+- **Date**: 2026-03-30
+- **Change**: `echo always > /sys/kernel/mm/transparent_hugepage/enabled`
+- **Applied via**: `sudo make optimize-power` (volatile — add `transparent_hugepage=always` to kernel cmdline for persistence)
+- **Notes**: Reduces TLB misses for mmap'd model files. Hard to isolate impact.
+- **Verdict**: KEEP — low risk
 
 ### 2.4 RADV_PERFTEST=nogttspill
 
-- **Date**: PENDING
-- **Change**: `export RADV_PERFTEST=nogttspill`
-- **Expected**: Fix pp degradation on Vulkan RADV (community-reported fix for Strix Halo)
-- **Benchmark**: Not yet run — needs Vulkan-specific benchmark comparison
-- **Verdict**: PENDING
+- **Date**: 2026-03-30
+- **Change**: `RADV_PERFTEST=nogttspill` persisted to `/etc/environment.d/radv-llm.conf`
+- **Applied via**: `sudo make optimize-power`
+- **Notes**: Prevents GTT spill management overhead on unified memory Vulkan. Takes effect on next login. For current session: `export RADV_PERFTEST=nogttspill`
+- **Verdict**: KEEP — persists across reboots
 
 ### 2.5 amdgpu.noretry=0
 
@@ -84,11 +88,19 @@ Living document tracking what was applied, tested, and the actual results. Each
 
 ### 3.1 KV Cache Quantization
 
-- **Date**: PENDING (sweep running)
-- **Change**: `-ctk q8_0 -ctv q8_0` / `-ctk q4_0 -ctv q4_0`
-- **Benchmark**: `data/benchmarks/kv-sweep-128k-*` (in progress)
-- **Expected**: Q8_0: ~50% less KV memory, negligible quality loss. Q4_0: ~75% less, noticeable quality impact.
-- **Verdict**: PENDING
+- **Date**: 2026-03-27
+- **Change**: `--kv-types f16,q8_0,q4_0` sweep
+- **Benchmark**: `data/benchmarks/kv-sweep-256k-*`
+- **Result** (Vulkan RADV, Qwen3.5-35B-A3B Q8, pp2048/tg1024):
+  - f16: 456 pp, 39.8 tg
+  - q8_0: 418 pp, 38.5 tg (slight Vulkan regression — unexpected)
+  - **q4_0: 460 pp, 41.1 tg** (fastest overall, +3% tg over f16)
+- **Result** (ROCm, same model):
+  - f16: 445 pp, 21.5 tg
+  - q8_0: 495 pp, 21.7 tg (+11% pp, same tg)
+  - q4_0: 494 pp, 21.8 tg (+11% pp, same tg)
+- **Conclusion**: q4_0 is the sweet spot on Vulkan (fastest tg + 75% less KV memory). On ROCm, KV quant helps pp but not tg.
+- **Verdict**: KEEP — use q4_0 KV as default for serving
 
 ### 3.2 MoE Batch Size `-b 256`
 
@@ -161,6 +173,24 @@ Living document tracking what was applied, tested, and the actual results. Each
 
 ---
 
+## Model Quant Shootout
+
+### Qwen3.5-35B-A3B — Q4_K_L vs Q4_K_XL vs Q8 (2026-03-30)
+
+- **Benchmark**: `data/benchmarks/qwen35-shootout-v2-*`
+- **Config**: Vulkan RADV, q4_0 KV cache, pp2048/tg1024, 2 reps
+- **RyzenAdj**: STAPM=81W (sustained ~70W due to HP firmware cap)
+
+| Quant | File Size | pp2048 (t/s) | tg1024 (t/s) | Recommendation |
+|-------|-----------|-------------|-------------|----------------|
+| UD-Q4_K_L | 18.8 GB | 825 | **57.0** | Fastest. Good quality. |
+| **UD-Q4_K_XL** | 20.7 GB | 835 | **56.4** | **Daily driver** — best quality/speed. |
+| Q8_0 | 34.4 GB | 850 | 51.4 | Best quality, 10% slower tg. |
+
+**Decision**: Keep UD-Q4_K_XL (daily driver) and Q8_0 (quality fallback). Q4_K_L can be deleted — Q4_K_XL is strictly better at only +2 GB.
+
+---
+
 ## How to Add Entries
 
 When testing a new optimization:
diff --git a/scripts/optimize/power-profile.sh b/scripts/optimize/power-profile.sh
new file mode 100644
index 0000000..2f5af4a
--- /dev/null
+++ b/scripts/optimize/power-profile.sh
@@ -0,0 +1,111 @@
+#!/usr/bin/env bash
+# Apply power profile and system tuning for LLM inference workloads
+# Requires root. Settings are volatile — use the systemd service for persistence.
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+source "$SCRIPT_DIR/../../lib/common.sh"
+source "$SCRIPT_DIR/../../lib/format.sh"
+
+require_root
+
+# ── Power limits via ryzenadj ─────────────────────────────
+STAPM=85000
+FAST=85000
+SLOW=85000
+APU_SLOW=85000
+
+if is_cmd ryzenadj; then
+    log_header "Power Profile (ryzenadj)"
+    log_info "Setting STAPM=${STAPM}mW, Fast=${FAST}mW, Slow=${SLOW}mW, APU=${APU_SLOW}mW"
+    ryzenadj \
+        --stapm-limit=$STAPM \
+        --fast-limit=$FAST \
+        --slow-limit=$SLOW \
+        --apu-slow-limit=$APU_SLOW 2>&1 | grep -E 'Successfully|Error|not supported' || true
+
+    # Verify what actually took effect
+    log_info "Verifying limits..."
+    ryzenadj -i 2>&1 | grep -E 'LIMIT|VALUE' | head -8
+    echo ""
+    log_warn "Note: HP firmware may cap PPT SLOW/APU at 70W regardless of setting"
+else
+    log_error "ryzenadj not found. Install: cd /tmp && git clone https://github.com/FlyGoat/RyzenAdj.git && cd RyzenAdj && mkdir build && cd build && cmake .. && make && sudo cp ryzenadj /usr/local/bin/"
+    exit 1
+fi
+
+# ── VM sysctl tuning ──────────────────────────────────────
+log_header "VM Sysctl Tuning"
+
+declare -A SYSCTLS=(
+    [vm.swappiness]=1
+    [vm.dirty_ratio]=40
+    [vm.dirty_background_ratio]=10
+    [vm.max_map_count]=500000
+    [vm.zone_reclaim_mode]=0
+)
+
+for KEY in "${!SYSCTLS[@]}"; do
+    VAL="${SYSCTLS[$KEY]}"
+    CURRENT=$(sysctl -n "$KEY" 2>/dev/null || echo "?")
+    if [[ "$CURRENT" == "$VAL" ]]; then
+        log_success "$KEY = $VAL (already set)"
+    else
+        sysctl -w "$KEY=$VAL" > /dev/null 2>&1
+        log_success "$KEY = $VAL (was $CURRENT)"
+    fi
+done
+
+# Persist sysctl settings
+SYSCTL_CONF="/etc/sysctl.d/99-llm-inference.conf"
+if [[ ! -f "$SYSCTL_CONF" ]]; then
+    log_info "Persisting to $SYSCTL_CONF"
+    cat > "$SYSCTL_CONF" << 'EOF'
+# LLM inference optimizations
+vm.swappiness = 1
+vm.dirty_ratio = 40
+vm.dirty_background_ratio = 10
+vm.max_map_count = 500000
+vm.zone_reclaim_mode = 0
+EOF
+    log_success "Sysctl config saved (persists across reboots)"
+else
+    log_info "Sysctl config already exists at $SYSCTL_CONF"
+fi
+
+# ── Transparent Huge Pages ────────────────────────────────
+log_header "Transparent Huge Pages"
+
+THP_ENABLED=$(cat /sys/kernel/mm/transparent_hugepage/enabled 2>/dev/null || echo "unknown")
+if [[ "$THP_ENABLED" == *"[always]"* ]]; then
+    log_success "THP = always (already set)"
+else
+    echo always > /sys/kernel/mm/transparent_hugepage/enabled 2>/dev/null || true
+    echo defer+madvise > /sys/kernel/mm/transparent_hugepage/defrag 2>/dev/null || true
+    log_success "THP = always, defrag = defer+madvise"
+fi
+log_info "For persistence, add to kernel cmdline: transparent_hugepage=always"
+
+# ── RADV nogttspill ───────────────────────────────────────
+log_header "Vulkan RADV Environment"
+
+RADV_CONF="/etc/environment.d/radv-llm.conf"
+if [[ ! -f "$RADV_CONF" ]]; then
+    mkdir -p /etc/environment.d
+    echo 'RADV_PERFTEST=nogttspill' > "$RADV_CONF"
+    log_success "RADV_PERFTEST=nogttspill persisted to $RADV_CONF"
+    log_info "Takes effect on next login. For this session: export RADV_PERFTEST=nogttspill"
+else
+    log_success "RADV config already exists at $RADV_CONF"
+fi
+
+# ── Summary ───────────────────────────────────────────────
+log_header "Phase 2 Optimization Summary"
+log_success "Power profile: ryzenadj limits applied (volatile — resets on reboot)"
+log_success "VM tuning: sysctl applied and persisted"
+log_success "THP: enabled (volatile — add to kernel cmdline for persistence)"
+log_success "RADV: nogttspill persisted"
+echo ""
+log_info "To persist ryzenadj across reboots:"
+log_info "  sudo cp $SCRIPT_DIR/../../configs/ryzenadj-llm.service /etc/systemd/system/"
+log_info "  sudo systemctl enable --now ryzenadj-llm.service"