diff --git a/scripts/make-test-fixtures.py b/scripts/make-test-fixtures.py
index e61ecf3..953f437 100755
--- a/scripts/make-test-fixtures.py
+++ b/scripts/make-test-fixtures.py
@@ -229,11 +229,89 @@ def build_gemma4_kv_shared(out, vestigial):
     return len(w)
 
 
+def build_moe_nested(out):
+    """A Qwen3-style MoE: per-expert tensors, nested config, decoy vision count.
+
+    Covers the MoE *load* path, which nothing else in CI touches — no MoE model is in
+    the matrix at all. Specifically: per-expert `mlp.experts.N.*` tensors being stacked
+    into `switch_mlp` by sanitize, and an expert count nested under `text_config`
+    rather than at top level.
+
+    What it does NOT cover, despite the nesting, is the #112 detection bug itself.
+    That bug was `num_local_experts` being the only spelling checked, so a config
+    declaring `num_experts` was called dense. But `modelTypeImpliesMoE` treats any
+    model_type containing "moe" as MoE regardless of keys, so `qwen3_moe` is caught by
+    that fallback whatever the config says — verified by reverting detection to the
+    single-key form and watching this fixture still enable streaming.
+
+    Reproducing #112 end to end needs a MoE architecture whose model_type does not
+    contain "moe" — deepseek_v3 (`n_routed_experts`) is the candidate. Its MLA
+    attention makes that a larger fixture; the profiler itself is covered directly by
+    ModelProfilerMoEDetectionTests in the meantime.
+    """
+    H, L, HEADS, KVH, HD = 64, 2, 4, 2, 16
+    INTER, MOE_INTER, EXPERTS, TOPK = 128, 32, 4, 2
+    rng = np.random.default_rng(0)
+
+    text_config = {
+        "model_type": "qwen3_moe",
+        "hidden_size": H,
+        "num_hidden_layers": L,
+        "intermediate_size": INTER,
+        "num_attention_heads": HEADS,
+        "num_key_value_heads": KVH,
+        "head_dim": HD,
+        "num_experts": EXPERTS,
+        "num_experts_per_tok": TOPK,
+        "moe_intermediate_size": MOE_INTER,
+        "decoder_sparse_step": 1,
+        "mlp_only_layers": [],
+        "rms_norm_eps": 1e-6,
+        "vocab_size": VOCAB,
+        "rope_theta": 10000.0,
+        "tie_word_embeddings": False,
+        "max_position_embeddings": 512,
+        "norm_topk_prob": True,
+    }
+    cfg = dict(text_config)
+    cfg["text_config"] = text_config
+    # Must not be mistaken for the language model's expert count.
+    cfg["vision_config"] = {"model_type": "qwen3_vl", "num_experts": 999}
+    json.dump(cfg, open(os.path.join(out, "config.json"), "w"), indent=2)
+
+    w = {
+        "model.embed_tokens.weight": rand(rng, VOCAB, H),
+        "model.norm.weight": ones(H),
+        "lm_head.weight": rand(rng, VOCAB, H),
+    }
+    for i in range(L):
+        p_ = f"model.layers.{i}"
+        w[f"{p_}.self_attn.q_proj.weight"] = rand(rng, HEADS * HD, H)
+        w[f"{p_}.self_attn.k_proj.weight"] = rand(rng, KVH * HD, H)
+        w[f"{p_}.self_attn.v_proj.weight"] = rand(rng, KVH * HD, H)
+        w[f"{p_}.self_attn.o_proj.weight"] = rand(rng, H, HEADS * HD)
+        w[f"{p_}.self_attn.q_norm.weight"] = ones(HD)
+        w[f"{p_}.self_attn.k_norm.weight"] = ones(HD)
+        w[f"{p_}.input_layernorm.weight"] = ones(H)
+        w[f"{p_}.post_attention_layernorm.weight"] = ones(H)
+        w[f"{p_}.mlp.gate.weight"] = rand(rng, EXPERTS, H)
+        # Per-expert tensors, the layout a real checkpoint ships; sanitize stacks
+        # these into switch_mlp.
+        for e in range(EXPERTS):
+            w[f"{p_}.mlp.experts.{e}.gate_proj.weight"] = rand(rng, MOE_INTER, H)
+            w[f"{p_}.mlp.experts.{e}.up_proj.weight"] = rand(rng, MOE_INTER, H)
+            w[f"{p_}.mlp.experts.{e}.down_proj.weight"] = rand(rng, H, MOE_INTER)
+
+    save_file(w, os.path.join(out, "model.safetensors"), metadata={"format": "pt"})
+    return len(w)
+
+
 FIXTURES = {
     "dense": (build_dense, {}),
     "stray-shard": (build_dense, {"stray_shard": True}),
     "kv-shared-absent": (build_gemma4_kv_shared, {"vestigial": False}),
     "kv-shared-present": (build_gemma4_kv_shared, {"vestigial": True}),
+    "moe-nested": (build_moe_nested, {}),
 }
 
 if __name__ == "__main__":
diff --git a/tests/fixtures/moe-nested/config.json b/tests/fixtures/moe-nested/config.json
new file mode 100644
index 0000000..bfee7eb
--- /dev/null
+++ b/tests/fixtures/moe-nested/config.json
@@ -0,0 +1,44 @@
+{
+  "model_type": "qwen3_moe",
+  "hidden_size": 64,
+  "num_hidden_layers": 2,
+  "intermediate_size": 128,
+  "num_attention_heads": 4,
+  "num_key_value_heads": 2,
+  "head_dim": 16,
+  "num_experts": 4,
+  "num_experts_per_tok": 2,
+  "moe_intermediate_size": 32,
+  "decoder_sparse_step": 1,
+  "mlp_only_layers": [],
+  "rms_norm_eps": 1e-06,
+  "vocab_size": 288,
+  "rope_theta": 10000.0,
+  "tie_word_embeddings": false,
+  "max_position_embeddings": 512,
+  "norm_topk_prob": true,
+  "text_config": {
+    "model_type": "qwen3_moe",
+    "hidden_size": 64,
+    "num_hidden_layers": 2,
+    "intermediate_size": 128,
+    "num_attention_heads": 4,
+    "num_key_value_heads": 2,
+    "head_dim": 16,
+    "num_experts": 4,
+    "num_experts_per_tok": 2,
+    "moe_intermediate_size": 32,
+    "decoder_sparse_step": 1,
+    "mlp_only_layers": [],
+    "rms_norm_eps": 1e-06,
+    "vocab_size": 288,
+    "rope_theta": 10000.0,
+    "tie_word_embeddings": false,
+    "max_position_embeddings": 512,
+    "norm_topk_prob": true
+  },
+  "vision_config": {
+    "model_type": "qwen3_vl",
+    "num_experts": 999
+  }
+}
\ No newline at end of file
diff --git a/tests/fixtures/moe-nested/model.safetensors b/tests/fixtures/moe-nested/model.safetensors
new file mode 100644
index 0000000..fab58e1
Binary files /dev/null and b/tests/fixtures/moe-nested/model.safetensors differ
diff --git a/tests/fixtures/moe-nested/tokenizer.json b/tests/fixtures/moe-nested/tokenizer.json
new file mode 100644
index 0000000..21c0e13
--- /dev/null
+++ b/tests/fixtures/moe-nested/tokenizer.json
@@ -0,0 +1,380 @@
+{
+  "version": "1.0",
+  "truncation": null,
+  "padding": null,
+  "added_tokens": [
+    {
+      "id": 0,
+      "content": "<|endoftext|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 1,
+      "content": "<|im_start|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 2,
+      "content": "<|im_end|>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    },
+    {
+      "id": 3,
+      "content": "<pad>",
+      "single_word": false,
+      "lstrip": false,
+      "rstrip": false,
+      "normalized": false,
+      "special": true
+    }
+  ],
+  "normalizer": null,
+  "pre_tokenizer": {
+    "type": "ByteLevel",
+    "add_prefix_space": false,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "post_processor": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": false,
+    "use_regex": true
+  },
+  "decoder": {
+    "type": "ByteLevel",
+    "add_prefix_space": true,
+    "trim_offsets": true,
+    "use_regex": true
+  },
+  "model": {
+    "type": "BPE",
+    "dropout": null,
+    "unk_token": null,
+    "continuing_subword_prefix": null,
+    "end_of_word_suffix": null,
+    "fuse_unk": false,
+    "byte_fallback": false,
+    "ignore_merges": false,
+    "vocab": {
+      "<|endoftext|>": 0,
+      "<|im_start|>": 1,
+      "<|im_end|>": 2,
+      "<pad>": 3,
+      "!": 4,
+      "\"": 5,
+      "#": 6,
+      "$": 7,
+      "%": 8,
+      "&": 9,
+      "'": 10,
+      "(": 11,
+      ")": 12,
+      "*": 13,
+      "+": 14,
+      ",": 15,
+      "-": 16,
+      ".": 17,
+      "/": 18,
+      "0": 19,
+      "1": 20,
+      "2": 21,
+      "3": 22,
+      "4": 23,
+      "5": 24,
+      "6": 25,
+      "7": 26,
+      "8": 27,
+      "9": 28,
+      ":": 29,
+      ";": 30,
+      "<": 31,
+      "=": 32,
+      ">": 33,
+      "?": 34,
+      "@": 35,
+      "A": 36,
+      "B": 37,
+      "C": 38,
+      "D": 39,
+      "E": 40,
+      "F": 41,
+      "G": 42,
+      "H": 43,
+      "I": 44,
+      "J": 45,
+      "K": 46,
+      "L": 47,
+      "M": 48,
+      "N": 49,
+      "O": 50,
+      "P": 51,
+      "Q": 52,
+      "R": 53,
+      "S": 54,
+      "T": 55,
+      "U": 56,
+      "V": 57,
+      "W": 58,
+      "X": 59,
+      "Y": 60,
+      "Z": 61,
+      "[": 62,
+      "\\": 63,
+      "]": 64,
+      "^": 65,
+      "_": 66,
+      "`": 67,
+      "a": 68,
+      "b": 69,
+      "c": 70,
+      "d": 71,
+      "e": 72,
+      "f": 73,
+      "g": 74,
+      "h": 75,
+      "i": 76,
+      "j": 77,
+      "k": 78,
+      "l": 79,
+      "m": 80,
+      "n": 81,
+      "o": 82,
+      "p": 83,
+      "q": 84,
+      "r": 85,
+      "s": 86,
+      "t": 87,
+      "u": 88,
+      "v": 89,
+      "w": 90,
+      "x": 91,
+      "y": 92,
+      "z": 93,
+      "{": 94,
+      "|": 95,
+      "}": 96,
+      "~": 97,
+      "¡": 98,
+      "¢": 99,
+      "£": 100,
+      "¤": 101,
+      "¥": 102,
+      "¦": 103,
+      "§": 104,
+      "¨": 105,
+      "©": 106,
+      "ª": 107,
+      "«": 108,
+      "¬": 109,
+      "®": 110,
+      "¯": 111,
+      "°": 112,
+      "±": 113,
+      "²": 114,
+      "³": 115,
+      "´": 116,
+      "µ": 117,
+      "¶": 118,
+      "·": 119,
+      "¸": 120,
+      "¹": 121,
+      "º": 122,
+      "»": 123,
+      "¼": 124,
+      "½": 125,
+      "¾": 126,
+      "¿": 127,
+      "À": 128,
+      "Á": 129,
+      "Â": 130,
+      "Ã": 131,
+      "Ä": 132,
+      "Å": 133,
+      "Æ": 134,
+      "Ç": 135,
+      "È": 136,
+      "É": 137,
+      "Ê": 138,
+      "Ë": 139,
+      "Ì": 140,
+      "Í": 141,
+      "Î": 142,
+      "Ï": 143,
+      "Ð": 144,
+      "Ñ": 145,
+      "Ò": 146,
+      "Ó": 147,
+      "Ô": 148,
+      "Õ": 149,
+      "Ö": 150,
+      "×": 151,
+      "Ø": 152,
+      "Ù": 153,
+      "Ú": 154,
+      "Û": 155,
+      "Ü": 156,
+      "Ý": 157,
+      "Þ": 158,
+      "ß": 159,
+      "à": 160,
+      "á": 161,
+      "â": 162,
+      "ã": 163,
+      "ä": 164,
+      "å": 165,
+      "æ": 166,
+      "ç": 167,
+      "è": 168,
+      "é": 169,
+      "ê": 170,
+      "ë": 171,
+      "ì": 172,
+      "í": 173,
+      "î": 174,
+      "ï": 175,
+      "ð": 176,
+      "ñ": 177,
+      "ò": 178,
+      "ó": 179,
+      "ô": 180,
+      "õ": 181,
+      "ö": 182,
+      "÷": 183,
+      "ø": 184,
+      "ù": 185,
+      "ú": 186,
+      "û": 187,
+      "ü": 188,
+      "ý": 189,
+      "þ": 190,
+      "ÿ": 191,
+      "Ā": 192,
+      "ā": 193,
+      "Ă": 194,
+      "ă": 195,
+      "Ą": 196,
+      "ą": 197,
+      "Ć": 198,
+      "ć": 199,
+      "Ĉ": 200,
+      "ĉ": 201,
+      "Ċ": 202,
+      "ċ": 203,
+      "Č": 204,
+      "č": 205,
+      "Ď": 206,
+      "ď": 207,
+      "Đ": 208,
+      "đ": 209,
+      "Ē": 210,
+      "ē": 211,
+      "Ĕ": 212,
+      "ĕ": 213,
+      "Ė": 214,
+      "ė": 215,
+      "Ę": 216,
+      "ę": 217,
+      "Ě": 218,
+      "ě": 219,
+      "Ĝ": 220,
+      "ĝ": 221,
+      "Ğ": 222,
+      "ğ": 223,
+      "Ġ": 224,
+      "ġ": 225,
+      "Ģ": 226,
+      "ģ": 227,
+      "Ĥ": 228,
+      "ĥ": 229,
+      "Ħ": 230,
+      "ħ": 231,
+      "Ĩ": 232,
+      "ĩ": 233,
+      "Ī": 234,
+      "ī": 235,
+      "Ĭ": 236,
+      "ĭ": 237,
+      "Į": 238,
+      "į": 239,
+      "İ": 240,
+      "ı": 241,
+      "IJ": 242,
+      "ij": 243,
+      "Ĵ": 244,
+      "ĵ": 245,
+      "Ķ": 246,
+      "ķ": 247,
+      "ĸ": 248,
+      "Ĺ": 249,
+      "ĺ": 250,
+      "Ļ": 251,
+      "ļ": 252,
+      "Ľ": 253,
+      "ľ": 254,
+      "Ŀ": 255,
+      "ŀ": 256,
+      "Ł": 257,
+      "ł": 258,
+      "Ń": 259,
+      "he": 260,
+      "ll": 261,
+      "hell": 262,
+      "te": 263,
+      "<|unused264|>": 264,
+      "<|unused265|>": 265,
+      "<|unused266|>": 266,
+      "<|unused267|>": 267,
+      "<|unused268|>": 268,
+      "<|unused269|>": 269,
+      "<|unused270|>": 270,
+      "<|unused271|>": 271,
+      "<|unused272|>": 272,
+      "<|unused273|>": 273,
+      "<|unused274|>": 274,
+      "<|unused275|>": 275,
+      "<|unused276|>": 276,
+      "<|unused277|>": 277,
+      "<|unused278|>": 278,
+      "<|unused279|>": 279,
+      "<|unused280|>": 280,
+      "<|unused281|>": 281,
+      "<|unused282|>": 282,
+      "<|unused283|>": 283,
+      "<|unused284|>": 284,
+      "<|unused285|>": 285,
+      "<|unused286|>": 286,
+      "<|unused287|>": 287
+    },
+    "merges": [
+      [
+        "h",
+        "e"
+      ],
+      [
+        "l",
+        "l"
+      ],
+      [
+        "he",
+        "ll"
+      ],
+      [
+        "t",
+        "e"
+      ]
+    ]
+  }
+}
\ No newline at end of file
diff --git a/tests/fixtures/moe-nested/tokenizer_config.json b/tests/fixtures/moe-nested/tokenizer_config.json
new file mode 100644
index 0000000..ee42569
--- /dev/null
+++ b/tests/fixtures/moe-nested/tokenizer_config.json
@@ -0,0 +1,8 @@
+{
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|im_end|>",
+  "pad_token": "<pad>",
+  "unk_token": "<|endoftext|>",
+  "chat_template": "{% for m in messages %}<|im_start|>{{ m['role'] }}\n{{ m['content'] }}<|im_end|>\n{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}
\ No newline at end of file
diff --git a/tests/test-fixtures.sh b/tests/test-fixtures.sh
index 230fa3a..92f2248 100755
--- a/tests/test-fixtures.sh
+++ b/tests/test-fixtures.sh
@@ -10,6 +10,9 @@
 #   stray-shard         #118: a .safetensors beside the index but absent from it
 #   kv-shared-absent    #120: gemma-4-e4b shape, shared layers ship no k/v
 #   kv-shared-present   b674:  gemma-4-e2b shape, shared layers ship k/v anyway
+#   moe-nested          MoE load path: per-expert tensors stacked into switch_mlp,
+#                       expert count nested under text_config. Note this does not
+#                       reproduce #112 — see scripts/make-test-fixtures.py for why.
 #
 # The output is gibberish by construction — the weights are random. A fixture passes
 # when the server loads it and produces *a* token, which is what exercises config
@@ -65,14 +68,21 @@ run_fixture() {
     local body
     body=$(curl -sf --max-time 60 "$url/v1/chat/completions" \
         -H 'Content-Type: application/json' \
-        -d '{"messages":[{"role":"user","content":"ping"}],"max_tokens":4,"stream":false}' 2>/dev/null)
+        -d '{"messages":[{"role":"user","content":"ping"}],"max_tokens":4,"temperature":0,"stream":false}' 2>/dev/null)
 
-    # Random weights make the text meaningless, so assert on the token count instead.
+    # What is being proved is that the checkpoint loaded and a forward pass ran, so the
+    # assertion is on prefill and a well-formed response — not on how many tokens came
+    # back. Random weights sample randomly: an early EOS yields zero completion tokens,
+    # which made an earlier version of this check pass or fail run to run. temperature 0
+    # removes the sampling variance; prompt_tokens is what actually evidences the prefill.
     if [ -n "$body" ] && echo "$body" | python3 -c '
 import json, sys
 d = json.load(sys.stdin)
-sys.exit(0 if d["usage"]["completion_tokens"] >= 1 else 1)' 2>/dev/null; then
-        pass "$name loaded and generated"
+ok = (d["usage"]["prompt_tokens"] >= 1
+      and isinstance(d["choices"][0]["message"]["content"], str)
+      and d["choices"][0].get("finish_reason"))
+sys.exit(0 if ok else 1)' 2>/dev/null; then
+        pass "$name loaded, prefilled and returned a well-formed completion"
     else
         fail "$name: no completion — $(echo "$body" | head -c 120)"
     fi
@@ -82,11 +92,37 @@ sys.exit(0 if d["usage"]["completion_tokens"] >= 1 else 1)' 2>/dev/null; then
 }
 
 log "Binary: $BINARY"
-for name in dense stray-shard kv-shared-absent kv-shared-present; do
+for name in dense stray-shard kv-shared-absent kv-shared-present moe-nested; do
     log "Shape: $name"
     run_fixture "$name"
 done
 
+# The MoE fixture again, this time with --stream-experts. A model classified dense
+# has the flag silently dropped and gets materialised whole, which is how #112
+# reached an OOM kill rather than an error message; asserting the flag was honoured
+# is the closest end-to-end check available.
+log "Shape: moe-nested (--stream-experts honoured)"
+MOE_LOG="/tmp/SwiftLM-test-fixture-moe-stream.log"
+"$BINARY" --model "$FIXTURE_DIR/moe-nested" --port "$PORT" --host "$HOST" \
+    --stream-experts > "$MOE_LOG" 2>&1 &
+SERVER_PID=$!
+ready=0
+for _ in $(seq 1 60); do
+    curl -sf "http://$HOST:$PORT/health" >/dev/null 2>&1 && { ready=1; break; }
+    kill -0 "$SERVER_PID" 2>/dev/null || break
+    sleep 1
+done
+if [ "$ready" -ne 1 ]; then
+    fail "moe-nested did not start with --stream-experts"
+elif grep -qi "is not MoE" "$MOE_LOG"; then
+    fail "moe-nested was classified dense: $(grep -i 'is not MoE' "$MOE_LOG" | head -1)"
+elif grep -q "SSD Expert Streaming enabled" "$MOE_LOG"; then
+    pass "moe-nested detected as MoE, --stream-experts honoured"
+else
+    fail "moe-nested: no streaming confirmation in log"
+fi
+cleanup
+
 log "═══════════════════════════════════════"
 log "Results: $PASS passed, $FAIL failed"
 log "═══════════════════════════════════════"