diff --git a/pyproject.toml b/pyproject.toml
index b388007..c0eacd9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hedit"
-version = "0.7.6a2"
+version = "0.7.6.dev4"
 description = "Multi-agent system for HED annotation generation and validation"
 readme = "PKG_README.md"
 requires-python = ">=3.12"
diff --git a/src/api/main.py b/src/api/main.py
index e1a9255..812e3d1 100644
--- a/src/api/main.py
+++ b/src/api/main.py
@@ -113,8 +113,8 @@ def create_openrouter_workflow(
     # Apply defaults from environment
     default_annotation_model = os.getenv("ANNOTATION_MODEL", "anthropic/claude-haiku-4.5")
     default_annotation_provider = os.getenv("ANNOTATION_PROVIDER", "anthropic")
-    default_eval_model = os.getenv("EVALUATION_MODEL", "openai/gpt-oss-120b")
-    default_eval_provider = os.getenv("EVALUATION_PROVIDER", "groq")
+    default_eval_model = os.getenv("EVALUATION_MODEL", "qwen/qwen3.5-397b-a17b")
+    default_eval_provider = os.getenv("EVALUATION_PROVIDER")
 
     # Resolve final values: parameter > env var > default
     actual_annotation_model = get_model_name(annotation_model or default_annotation_model)
@@ -240,8 +240,8 @@ def create_byok_vision_agent(
         Configured VisionAgent using the user's key and model settings
     """
     # Use user-provided settings or fall back to server defaults
-    default_vision_model = os.getenv("VISION_MODEL", "qwen/qwen3-vl-30b-a3b-instruct")
-    default_vision_provider = os.getenv("VISION_PROVIDER", "deepinfra/fp8")
+    default_vision_model = os.getenv("VISION_MODEL", "qwen/qwen3-vl-32b-instruct")
+    default_vision_provider = os.getenv("VISION_PROVIDER", "novita")
 
     actual_model = vision_model if vision_model else default_vision_model
     actual_temperature = temperature if temperature is not None else 0.3
@@ -354,7 +354,7 @@ def get_default_path(docker_path: str, local_path: str) -> str | None:
         # Log configuration (env vars are read by create_openrouter_workflow)
         print("Using OpenRouter with models:")
         print(f"  Annotation: {os.getenv('ANNOTATION_MODEL', 'anthropic/claude-haiku-4.5')}")
-        print(f"  Evaluation: {os.getenv('EVALUATION_MODEL', 'qwen/qwen3-235b-a22b-2507')}")
+        print(f"  Evaluation: {os.getenv('EVALUATION_MODEL', 'qwen/qwen3.5-397b-a17b')}")
         print(f"  Provider (annotation): {os.getenv('ANNOTATION_PROVIDER', 'anthropic')}")
         print(f"  Provider (eval): {os.getenv('EVALUATION_PROVIDER', '') or '(auto-routed)'}")
 
@@ -395,8 +395,8 @@ def get_default_path(docker_path: str, local_path: str) -> str | None:
 
     # Initialize vision agent (only for OpenRouter)
     if llm_provider == "openrouter":
-        vision_model = os.getenv("VISION_MODEL", "qwen/qwen3-vl-30b-a3b-instruct")
-        vision_provider = os.getenv("VISION_PROVIDER", "deepinfra/fp8")
+        vision_model = os.getenv("VISION_MODEL", "qwen/qwen3-vl-32b-instruct")
+        vision_provider = os.getenv("VISION_PROVIDER", "novita")
 
         print(f"Initializing vision model: {vision_model} (provider: {vision_provider})")
 
@@ -666,7 +666,7 @@ async def annotate(
             model_name = (
                 request.model
                 or req.headers.get("x-openrouter-model")
-                or os.getenv("ANNOTATION_MODEL", "openai/gpt-oss-120b")
+                or os.getenv("ANNOTATION_MODEL", "anthropic/claude-haiku-4.5")
             )
             temperature = request.temperature
             if temperature is None:
@@ -815,7 +815,7 @@ async def annotate_from_image(
                 validator_path=_byok_config.get("validator_path"),
                 use_js_validator=_byok_config.get("use_js_validator", True),
             )
-            # Note: Vision agent uses its own provider (deepinfra/fp8 for qwen-vl)
+            # Note: Vision agent uses its own provider (novita for qwen-vl)
             # Only pass provider_override to vision if a custom vision_model was specified
             vision_provider = provider_override if vision_model_override else None
             active_vision_agent = create_byok_vision_agent(
@@ -877,7 +877,7 @@ async def annotate_from_image(
             model_name = (
                 request.model
                 or req.headers.get("x-openrouter-model")
-                or os.getenv("ANNOTATION_MODEL", "openai/gpt-oss-120b")
+                or os.getenv("ANNOTATION_MODEL", "anthropic/claude-haiku-4.5")
             )
             temperature = request.temperature
             if temperature is None:
@@ -967,7 +967,7 @@ async def _collect_stream_telemetry(
     model_name = (
         request.model
         or req.headers.get("x-openrouter-model")
-        or os.getenv("ANNOTATION_MODEL", "openai/gpt-oss-120b")
+        or os.getenv("ANNOTATION_MODEL", "anthropic/claude-haiku-4.5")
     )
     temperature = request.temperature
     if temperature is None:
@@ -1386,7 +1386,7 @@ async def annotate_from_image_stream(
                 validator_path=_byok_config.get("validator_path"),
                 use_js_validator=_byok_config.get("use_js_validator", True),
             )
-            # Note: Vision agent uses its own provider (deepinfra/fp8 for qwen-vl)
+            # Note: Vision agent uses its own provider (novita for qwen-vl)
             # Only pass provider_override to vision if a custom vision_model was specified
             vision_provider = provider_override if vision_model_override else None
             active_vision_agent = create_byok_vision_agent(
@@ -1785,7 +1785,7 @@ async def submit_feedback(request: FeedbackRequest) -> FeedbackResponse:
                 )
 
                 # Create LLM for triage
-                model = os.getenv("ANNOTATION_MODEL", "openai/gpt-oss-120b")
+                model = os.getenv("ANNOTATION_MODEL", "anthropic/claude-haiku-4.5")
                 provider = os.getenv("LLM_PROVIDER_PREFERENCE", "")
                 llm = create_openrouter_llm(
                     model=model,
diff --git a/src/api/models.py b/src/api/models.py
index 4f213a1..5d78102 100644
--- a/src/api/models.py
+++ b/src/api/models.py
@@ -41,7 +41,7 @@ class AnnotationRequest(BaseModel):
     model: str | None = Field(
         default=None,
         description="Override model for annotation (BYOK mode only, e.g., 'openai/gpt-4o')",
-        examples=["openai/gpt-4o", "anthropic/claude-3.5-sonnet", "qwen/qwen3-235b-a22b-2507"],
+        examples=["anthropic/claude-haiku-4.5", "qwen/qwen3.5-122b-a10b", "qwen/qwen3.5-397b-a17b"],
     )
     provider: str | None = Field(
         default=None,
@@ -177,7 +177,7 @@ class ImageAnnotationRequest(BaseModel):
     vision_model: str | None = Field(
         default=None,
         description="Override vision model for image description (BYOK mode only)",
-        examples=["qwen/qwen3-vl-30b-a3b-instruct", "openai/gpt-4o"],
+        examples=["qwen/qwen3-vl-32b-instruct", "qwen/qwen3-vl-235b-a22b-instruct"],
     )
     provider: str | None = Field(
         default=None,
diff --git a/src/cli/config.py b/src/cli/config.py
index 12ef5f4..c1e1fb5 100644
--- a/src/cli/config.py
+++ b/src/cli/config.py
@@ -41,13 +41,13 @@
 DEFAULT_MODEL = "anthropic/claude-haiku-4.5"
 DEFAULT_PROVIDER = "anthropic"
 
-# Evaluation model: Qwen3-235B (auto-routed by OpenRouter)
-DEFAULT_EVAL_MODEL = "qwen/qwen3-235b-a22b-2507"
+# Evaluation model: Qwen3.5-397B MoE (most capable Qwen, high throughput)
+DEFAULT_EVAL_MODEL = "qwen/qwen3.5-397b-a17b"
 DEFAULT_EVAL_PROVIDER = None
 
-# Vision model: Qwen3-VL for image descriptions
-DEFAULT_VISION_MODEL = "qwen/qwen3-vl-30b-a3b-instruct"
-DEFAULT_VISION_PROVIDER = "deepinfra/fp8"
+# Vision model: Qwen3-VL-32B for image descriptions
+DEFAULT_VISION_MODEL = "qwen/qwen3-vl-32b-instruct"
+DEFAULT_VISION_PROVIDER = None
 
 
 class CredentialsConfig(BaseModel):
diff --git a/src/cli/local_executor.py b/src/cli/local_executor.py
index e7bf48f..65e6853 100644
--- a/src/cli/local_executor.py
+++ b/src/cli/local_executor.py
@@ -77,11 +77,10 @@ def __init__(
 
         Args:
             api_key: OpenRouter API key (required for LLM operations, optional for health/validate)
-            model: Model for text annotation (default: openai/gpt-oss-120b)
-            eval_model: Model for evaluation/assessment agents (default: same as model).
-                       Use a consistent model like qwen/qwen3-235b-a22b for fair benchmarking.
-            eval_provider: Provider for evaluation model (e.g., Cerebras for qwen models)
-            vision_model: Model for image annotation (default: qwen/qwen3-vl-30b-a3b-instruct)
+            model: Model for text annotation (default: anthropic/claude-haiku-4.5)
+            eval_model: Model for evaluation/assessment agents (default: qwen/qwen3.5-397b-a17b)
+            eval_provider: Provider for evaluation model (None = OpenRouter auto-routes)
+            vision_model: Model for image annotation (default: qwen/qwen3-vl-32b-instruct)
             provider: Provider preference (cleared if custom model specified)
             temperature: LLM temperature (0.0-1.0)
             schema_dir: Optional directory with JSON schemas (None = fetch from GitHub)
diff --git a/src/cli/main.py b/src/cli/main.py
index 13af76a..ec04943 100644
--- a/src/cli/main.py
+++ b/src/cli/main.py
@@ -102,7 +102,7 @@
     typer.Option(
         "--model",
         "-m",
-        help="Model to use (e.g., openai/gpt-oss-120b, gpt-4o-mini)",
+        help="Model to use (e.g., anthropic/claude-haiku-4.5, qwen/qwen3.5-122b-a10b)",
     ),
 ]
 
diff --git a/src/scripts/process_feedback.py b/src/scripts/process_feedback.py
index 6137834..1baa5d3 100644
--- a/src/scripts/process_feedback.py
+++ b/src/scripts/process_feedback.py
@@ -56,7 +56,7 @@ async def process_feedback_file(
         sys.exit(1)
 
     # Get model configuration from environment
-    model = os.getenv("ANNOTATION_MODEL", "openai/gpt-oss-120b")
+    model = os.getenv("ANNOTATION_MODEL", "anthropic/claude-haiku-4.5")
     provider = os.getenv("LLM_PROVIDER_PREFERENCE", "")
 
     # Create LLM
diff --git a/src/utils/litellm_llm.py b/src/utils/litellm_llm.py
index 5533f3c..f2a3246 100644
--- a/src/utils/litellm_llm.py
+++ b/src/utils/litellm_llm.py
@@ -29,7 +29,7 @@
 
 
 def create_litellm_openrouter(
-    model: str = "openai/gpt-oss-120b",
+    model: str = "anthropic/claude-haiku-4.5",
     api_key: str | None = None,
     temperature: float = 0.1,
     max_tokens: int | None = None,
diff --git a/src/utils/openrouter_llm.py b/src/utils/openrouter_llm.py
index b3a68b4..f0110fc 100644
--- a/src/utils/openrouter_llm.py
+++ b/src/utils/openrouter_llm.py
@@ -13,7 +13,7 @@
 
 
 def create_openrouter_llm(
-    model: str = "openai/gpt-oss-120b",
+    model: str = "anthropic/claude-haiku-4.5",
     api_key: str | None = None,
     temperature: float = 0.1,
     max_tokens: int | None = None,
@@ -28,7 +28,7 @@ def create_openrouter_llm(
     to include cache_control markers for 90% cost reduction on cache hits.
 
     Args:
-        model: Model identifier (e.g., "openai/gpt-oss-120b", "anthropic/claude-haiku-4.5")
+        model: Model identifier (e.g., "anthropic/claude-haiku-4.5", "qwen/qwen3.5-397b-a17b")
         api_key: OpenRouter API key (defaults to OPENROUTER_API_KEY env var)
         temperature: Sampling temperature (0.0-1.0)
         max_tokens: Maximum tokens to generate
diff --git a/src/version.py b/src/version.py
index a9076d1..3ae7347 100644
--- a/src/version.py
+++ b/src/version.py
@@ -1,7 +1,7 @@
 """Version information for HEDit."""
 
-__version__ = "0.7.6a2"
-__version_info__ = (0, 7, 6, "alpha")
+__version__ = "0.7.6.dev4"
+__version_info__ = (0, 7, 6, "dev")
 
 
 def get_version() -> str: