diff --git a/llama/models.ini b/llama/models.ini
index 2733479..7f3d06f 100644
--- a/llama/models.ini
+++ b/llama/models.ini
@@ -18,7 +18,6 @@ models-max    = 1
 ; Qwen3.6-35B-A3B — Alibaba, MoE 3B-active + multimodal. Fast all-rounder.
 ; ---------------------------------------------------------------------------
 [qwen3.6-35b-a3b]
-load-on-startup    = true
 model              = /home/spencer/.cache/models/Qwen3.6-35B-A3B-UD-Q8_K_XL.gguf
 mmproj             = /home/spencer/.cache/models/Qwen3.6-35B-A3B-mmproj-BF16.gguf
 spec-type          = draft-mtp
@@ -38,7 +37,7 @@ repeat-penalty     = 1.0
 ; Qwen3.8-27B — Alibaba, dense + multimodal. Deep reasoning.
 ; ---------------------------------------------------------------------------
 [qwen3.8-27b]
-model              = /home/spencer/.cache/models/Qwen3.8-27B-UD-Q8_K_XL.gguf
+model              = /home/spencer/.cache/models/Qwen3.8-27B-UD-Q4_K_M.gguf
 mmproj             = /home/spencer/.cache/models/Qwen3.8-27B-mmproj-BF16.gguf
 spec-type          = draft-mtp
 spec-draft-n-max   = 4
@@ -66,3 +65,27 @@ top-p              = 0.95
 top-k              = 40
 min-p              = 0.01
 repeat-penalty     = 1.0
+
+; ---------------------------------------------------------------------------
+; Qwen3.8-Flash-Next — Qwen4-preview arch (qwen4exp), 125B total / 6B active.
+; Thinking-mode sampling per official card. No MTP (unsupported yet); vision via
+; mmproj-BF16. UD-Q4_K_XL = ~104 GiB, PLE ngram table (~30 GiB)
+; lazily served from SSD via mmap → ~74 GiB resident, fits 124 GiB GTT solo.
+; Speculative decoding via ngram-mod (long-match proposal).
+; ---------------------------------------------------------------------------
+[qwen3.8-flash-next]
+model              = /home/spencer/.cache/models/Qwen3.8-Flash-Next-UD-Q4_K_XL-00001-of-00004.gguf
+mmproj             = /home/spencer/.cache/models/Qwen3.8-Flash-Next-mmproj-BF16.gguf
+chat-template-file = /home/spencer/.config/llama.cpp/qwen-fixed-chat_template.jinja
+reasoning-format   = deepseek
+reasoning          = auto
+reasoning-preserve = true
+spec-type          = ngram-mod
+spec-ngram-mod-n-max = 32
+spec-ngram-mod-n-min = 16
+temperature        = 1.0
+top-p              = 0.95
+top-k              = 20
+min-p              = 0.0
+presence-penalty   = 0.0
+repeat-penalty     = 1.0
diff --git a/pi/models.json b/pi/models.json
index 4b8a370..e0d2154 100644
--- a/pi/models.json
+++ b/pi/models.json
@@ -76,6 +76,26 @@
             "max": null
           }
         },
+        "qwen3.8-flash-next": {
+          "name": "Qwen3.8 Flash Next (llama.cpp)",
+          "reasoning": true,
+          "input": ["text", "image"],
+          "contextWindow": 262144,
+          "maxTokens": 262144,
+          "compat": {
+            "thinkingFormat": "reasoning_effort",
+            "supportsReasoningEffort": true
+          },
+          "thinkingLevelMap": {
+            "off": "none",
+            "minimal": null,
+            "low": "low",
+            "medium": "medium",
+            "high": null,
+            "xhigh": "xhigh",
+            "max": null
+          }
+        },
         "qwen3-coder-next-80b-a3b": {
           "name": "Qwen3 Coder Next 80B A3B (llama.cpp)",
           "reasoning": false,