diff --git a/docs/alora.md b/docs/alora.md
index 75d02037..63b4a98a 100644
--- a/docs/alora.md
+++ b/docs/alora.md
@@ -37,7 +37,7 @@ Use the `m alora train` command to fine-tune a LoRA or aLoRA adapter requirement
 
 ```bash
 m alora train path/to/data.jsonl \
-  --basemodel ibm-granite/granite-3.2-8b-instruct \
+  --basemodel ibm-granite/granite-4.0-micro \
   --outfile ./checkpoints/alora_adapter \
   --adapter alora \
   --epochs 6 \
@@ -47,6 +47,10 @@ m alora train path/to/data.jsonl \
   --grad-accum 4
 ```
 
+> **Note on Model Selection**: Use non-hybrid models (e.g., `granite-4.0-micro`) for aLoRA training.
+> Hybrid models (`granite-4.0-h-micro`) are recommended for general inference but adapters should be
+> trained on non-hybrid base models for compatibility with the `ibm-granite/rag-intrinsics-lib` repository.
+
 ### 📌 Parameters
 
 | Flag              | Type    | Default   | Description                                      |
diff --git a/docs/dev/requirement_aLoRA_rerouting.md b/docs/dev/requirement_aLoRA_rerouting.md
index 74794f8a..546f1606 100644
--- a/docs/dev/requirement_aLoRA_rerouting.md
+++ b/docs/dev/requirement_aLoRA_rerouting.md
@@ -37,10 +37,10 @@ from mellea.core import Requirement
 from mellea.backends.adapters import GraniteCommonAdapter
 
 m = start_session(
-    "huggingface.LocalHFBackend:ibm-granite/granite-3.2-8b-instruct")
+    "huggingface.LocalHFBackend:ibm-granite/granite-4.0-micro")
 
 # By default, the AloraRequirement uses a GraniteCommonAdapter with "requirement_check".
-m.backend.add_adapter(GraniteCommonAdapter("ibm-granite/rag-intrinsics-lib", "requirement_check", base_model_name="granite-3.2-8b-instruct"))
+m.backend.add_adapter(GraniteCommonAdapter("ibm-granite/rag-intrinsics-lib", "requirement_check", base_model_name="granite-4.0-micro"))
 
 m.instruct(
     "Corporate wants you to find the difference between these two strings:\n\naaa\naba")
diff --git a/docs/examples/aLora/101_example.py b/docs/examples/aLora/101_example.py
index 5ad625f9..1d5a6695 100644
--- a/docs/examples/aLora/101_example.py
+++ b/docs/examples/aLora/101_example.py
@@ -13,7 +13,7 @@
 
 # Define a backend and add the constraint aLora
 backend = LocalHFBackend(
-    model_id="ibm-granite/granite-3.2-8b-instruct", cache=SimpleLRUCache(5)
+    model_id="ibm-granite/granite-4.0-h-micro", cache=SimpleLRUCache(5)
 )
 
 custom_stembolt_failure_constraint = HFConstraintAlora(
diff --git a/docs/examples/generative_slots/generative_gsm8k.py b/docs/examples/generative_slots/generative_gsm8k.py
index efaf1e0d..44da109d 100644
--- a/docs/examples/generative_slots/generative_gsm8k.py
+++ b/docs/examples/generative_slots/generative_gsm8k.py
@@ -1,6 +1,6 @@
 # pytest: ollama, qualitative, llm, slow
 
-"""Example of chain-of-thought reasoning on a mathematical question from the GSM8K dataset, structured as code for improved performance with the granite-3.3-8B model. The original accuracy in standard "thinking" mode is approximately 80%, while this implementation achieves 85-89% accuracy—up to 9 points higher.
+"""Example of chain-of-thought reasoning on a mathematical question from the GSM8K dataset, structured as code for improved performance with Granite 4 models. The original accuracy in standard "thinking" mode is approximately 80%, while this implementation achieves 85-89% accuracy—up to 9 points higher.
 
 This demonstrates that generative decorators are sufficient for complex reasoning tasks: not only do they maintain or improve performance, but they also significantly enhance observability and control. For instance, the structured Thought titles can be easily surfaced in a UI, providing instant insight into the model's reasoning process.
 """
diff --git a/docs/examples/instruct_validate_repair/101_email_with_validate.py b/docs/examples/instruct_validate_repair/101_email_with_validate.py
index 742da99a..b510cf26 100644
--- a/docs/examples/instruct_validate_repair/101_email_with_validate.py
+++ b/docs/examples/instruct_validate_repair/101_email_with_validate.py
@@ -3,7 +3,7 @@
 from docs.examples.helper import req_print, w
 from mellea import start_session
 from mellea.backends import ModelOption
-from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.stdlib.sampling import RejectionSamplingStrategy
 
 # create a session using Granite 4 Micro (3B) on Ollama and a simple context [see below]
diff --git a/docs/examples/intrinsics/intrinsics.py b/docs/examples/intrinsics/intrinsics.py
index b7fcb267..862c8399 100644
--- a/docs/examples/intrinsics/intrinsics.py
+++ b/docs/examples/intrinsics/intrinsics.py
@@ -13,7 +13,7 @@
 # Create the backend. Example for a VLLM Server. Commented out in favor of the hugging face code for now.
 # # Assumes a locally running VLLM server.
 # backend = OpenAIBackend(
-#     model_id="ibm-granite/granite-3.3-8b-instruct",
+#     model_id="ibm-granite/granite-4.0-micro",
 #     base_url="http://0.0.0.0:8000/v1",
 #     api_key="EMPTY",
 # )
diff --git a/docs/examples/m_serve/client.py b/docs/examples/m_serve/client.py
index 31154be5..ea36c7c9 100644
--- a/docs/examples/m_serve/client.py
+++ b/docs/examples/m_serve/client.py
@@ -6,7 +6,7 @@
 
 response = client.chat.completions.create(
     messages=[{"role": "user", "content": "Find all the real roots of x^3 + 1."}],
-    model="granite3.3:8b",
+    model="granite4:micro-h",
 )
 
 print(response.choices[0])
diff --git a/docs/examples/tutorial/document_mobject.py b/docs/examples/tutorial/document_mobject.py
index d413d7d7..a3b55c50 100644
--- a/docs/examples/tutorial/document_mobject.py
+++ b/docs/examples/tutorial/document_mobject.py
@@ -1,7 +1,7 @@
 # pytest: ollama, qualitative, llm, requires_heavy_ram
 
 from mellea.backends import model_ids
-from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.stdlib.components.docs.richdocument import RichDocument
 
 rd = RichDocument.from_document_file("https://arxiv.org/pdf/1906.04043")
diff --git a/docs/kv_smash/hf_example.py b/docs/kv_smash/hf_example.py
index c5db6967..1964e213 100644
--- a/docs/kv_smash/hf_example.py
+++ b/docs/kv_smash/hf_example.py
@@ -2,7 +2,7 @@
 
 from mellea.backends import ModelOption
 from mellea.backends.huggingface import LocalHFBackend
-from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.core import CBlock
 from mellea.stdlib.components import Message
 from mellea.stdlib.context import ChatContext
@@ -30,7 +30,7 @@ async def example():
         role="user",
         content="What is the likely ZIP code of Nathan Fulton's work address?",
     )
-    backend = LocalHFBackend(model_id=IBM_GRANITE_3_3_8B)
+    backend = LocalHFBackend(model_id=IBM_GRANITE_4_HYBRID_MICRO)
     mot = await backend._generate_from_context_with_kv_cache(
         action=msg, ctx=ctx, model_options={ModelOption.MAX_NEW_TOKENS: 64}
     )
diff --git a/docs/kv_smash/kv_with_chat.py b/docs/kv_smash/kv_with_chat.py
index bdf6f38e..7f43c539 100644
--- a/docs/kv_smash/kv_with_chat.py
+++ b/docs/kv_smash/kv_with_chat.py
@@ -2,9 +2,9 @@
 
 from mellea.backends.huggingface import LocalHFBackend
 from mellea.backends.kv_block_helpers import DynamicCache, merge_dynamic_caches
-from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 
-backend = LocalHFBackend(model_id=IBM_GRANITE_3_3_8B)
+backend = LocalHFBackend(model_id=IBM_GRANITE_4_HYBRID_MICRO)
 
 model = backend._model
 tokenizer = backend._tokenizer
diff --git a/docs/kv_smash/kvcache.py b/docs/kv_smash/kvcache.py
index 51e1b5cc..40a472ac 100644
--- a/docs/kv_smash/kvcache.py
+++ b/docs/kv_smash/kvcache.py
@@ -9,7 +9,7 @@
 
 from mellea.backends.kv_block_helpers import DynamicCache, merge_dynamic_caches
 
-model_id = "ibm-granite/granite-3.3-8b-instruct"
+model_id = "ibm-granite/granite-4.0-tiny-preview"
 device = torch.device("mps")
 model = AutoModelForCausalLM.from_pretrained(model_id).to(device)
 tokenizer: PreTrainedTokenizer = AutoTokenizer.from_pretrained(model_id)
diff --git a/docs/tutorial.md b/docs/tutorial.md
index 1892c250..f88a8f89 100644
--- a/docs/tutorial.md
+++ b/docs/tutorial.md
@@ -791,7 +791,7 @@ We will train a lightweight adapter with the `m alora train` command on this sma
 ```bash
 m alora train /to/stembolts_data.jsonl \
   --promptfile ./prompt_config.json \
-  --basemodel ibm-granite/granite-3.2-8b-instruct \
+  --basemodel ibm-granite/granite-4.0-h-micro \
   --outfile ./checkpoints/alora_adapter \
   --adapter alora \
   --epochs 6 \
@@ -1321,9 +1321,9 @@ Assuming a component's TemplateRepresentation contains a `template_order` field,
 
 If the default formatter searches the template path or the package, it uses the following logic:
 - look in the `.../templates/prompts/...` directory
-- traverse sub-directories in that path that match the formatter's model id (ie `ibm-granite/granite-3.2-8b-instruct` will match `.../templates/prompts/granite/granite-3-2/instruct`) or default (ie `.../templates/prompts/default`)
+- traverse sub-directories in that path that match the formatter's model id (ie `ibm-granite/granite-4.0-h-micro` will match `.../templates/prompts/granite/granite-4-0-h/micro`) or default (ie `.../templates/prompts/default`)
 - return the template at the deepest directory path
-- the default template formatter assumes that a model will only have one match in any given directory; in other words, traversing a `templates` directory with both `prompts/granite/...` and `prompts/ibm/...` for `ibm-granite/granite-3.2-8b-instruct` should not happen
+- the default template formatter assumes that a model will only have one match in any given directory; in other words, traversing a `templates` directory with both `prompts/granite/...` and `prompts/ibm/...` for `ibm-granite/granite-4.0-h-micro` should not happen
 
 #### Editing an Existing Class
 To customize the template and template representation of an existing class, simply create a new class that inherits from the class you want to edit. Then, override the format_for_llm function and create a new template.
diff --git a/mellea/backends/model_ids.py b/mellea/backends/model_ids.py
index d20d3179..ed4621f1 100644
--- a/mellea/backends/model_ids.py
+++ b/mellea/backends/model_ids.py
@@ -25,14 +25,28 @@ class ModelIdentifier:
 #### IBM models ####
 ####################
 
-IBM_GRANITE_4_MICRO_3B = ModelIdentifier(
-    hf_model_name="ibm-granite/granite-4.0-micro",
-    ollama_name="granite4:micro",
+# Granite 4 Hybrid Models (Recommended for general use)
+IBM_GRANITE_4_HYBRID_MICRO = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.0-h-micro",
+    ollama_name="granite4:micro-h",
+    watsonx_name=None,  # Only h-small available on Watsonx
+)
+
+IBM_GRANITE_4_HYBRID_TINY = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.0-h-tiny",
+    ollama_name="granite4:tiny-h",
+    watsonx_name=None,  # Only h-small available on Watsonx
+)
+
+IBM_GRANITE_4_HYBRID_SMALL = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.0-h-small",
+    ollama_name="granite4:small-h",
     watsonx_name="ibm/granite-4-h-small",
 )
-# todo: watsonx model is different from ollama model - should be same.
 
 
+# Deprecated Granite 3 models - kept for backward compatibility
+# These maintain their original model references (not upgraded to Granite 4)
 IBM_GRANITE_3_2_8B = ModelIdentifier(
     hf_model_name="ibm-granite/granite-3.2-8b-instruct",
     ollama_name="granite3.2:8b",
@@ -45,6 +59,23 @@ class ModelIdentifier:
     watsonx_name="ibm/granite-3-3-8b-instruct",
 )
 
+# Deprecated: Use IBM_GRANITE_4_HYBRID_MICRO or IBM_GRANITE_4_HYBRID_SMALL instead
+# Kept for backward compatibility with per-backend model selection:
+# - Ollama/HF: Uses MICRO (fits in CI memory constraints)
+# - Watsonx: Uses SMALL (required for watsonx support)
+IBM_GRANITE_4_MICRO_3B = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-4.0-h-micro",
+    ollama_name="granite4:micro-h",
+    watsonx_name="ibm/granite-4-h-small",
+)
+
+# Granite 3.3 Vision Model (2B)
+IBM_GRANITE_3_3_VISION_2B = ModelIdentifier(
+    hf_model_name="ibm-granite/granite-vision-3.3-2b",
+    ollama_name="ibm/granite3.3-vision:2b",
+    watsonx_name=None,
+)
+
 IBM_GRANITE_GUARDIAN_3_0_2B = ModelIdentifier(
     hf_model_name="ibm-granite/granite-guardian-3.0-2b",
     ollama_name="granite3-guardian:2b",
diff --git a/mellea/backends/openai.py b/mellea/backends/openai.py
index 8ad7796e..b9ae9432 100644
--- a/mellea/backends/openai.py
+++ b/mellea/backends/openai.py
@@ -83,7 +83,7 @@ def __init__(
         """Initialize and OpenAI compatible backend. For any additional kwargs that you need to pass the the client, pass them as a part of **kwargs.
 
         Args:
-            model_id : A generic model identifier or OpenAI compatible string. Defaults to model_ids.IBM_GRANITE_3_3_8B.
+            model_id : A generic model identifier or OpenAI compatible string. Defaults to model_ids.IBM_GRANITE_4_HYBRID_MICRO.
             formatter: A custom formatter based on backend.If None, defaults to TemplateFormatter
             base_url : Base url for LLM API. Defaults to None.
             model_options : Generation options to pass to the LLM. Defaults to None.
diff --git a/mellea/backends/watsonx.py b/mellea/backends/watsonx.py
index 0fe1c2bd..8e03507f 100644
--- a/mellea/backends/watsonx.py
+++ b/mellea/backends/watsonx.py
@@ -54,7 +54,7 @@ class WatsonxAIBackend(FormatterBackend):
 
     def __init__(
         self,
-        model_id: str | ModelIdentifier = model_ids.IBM_GRANITE_3_3_8B,
+        model_id: str | ModelIdentifier = model_ids.IBM_GRANITE_4_HYBRID_SMALL,
         formatter: ChatFormatter | None = None,
         base_url: str | None = None,
         model_options: dict | None = None,
@@ -66,7 +66,7 @@ def __init__(
         """A generic watsonx backend that wraps around the ibm_watsonx_ai sdk.
 
         Args:
-            model_id  : Model id. Defaults to model_ids.IBM_GRANITE_3_3_8B.
+            model_id  : Model id. Defaults to model_ids.IBM_GRANITE_4_HYBRID_SMALL.
             formatter : input formatter. Defaults to TemplateFormatter in __init__.
             base_url  : url for watson ML deployment. Defaults to env(WATSONX_URL).
             model_options : Global model options to pass to the model. Defaults to None.
diff --git a/test/backends/test_huggingface.py b/test/backends/test_huggingface.py
index 6b5bfb85..a8ee0a48 100644
--- a/test/backends/test_huggingface.py
+++ b/test/backends/test_huggingface.py
@@ -27,7 +27,7 @@
 ]
 
 from mellea import MelleaSession
-from mellea.backends import ModelOption
+from mellea.backends import ModelOption, model_ids
 from mellea.backends.adapters import GraniteCommonAdapter
 from mellea.backends.cache import SimpleLRUCache
 from mellea.backends.huggingface import LocalHFBackend, _assert_correct_adapters
@@ -46,7 +46,12 @@
 
 @pytest.fixture(scope="module")
 def backend():
-    """Shared HuggingFace backend for all tests in this module."""
+    """Shared HuggingFace backend for all tests in this module.
+
+    Uses Granite 3.3-8b for aLoRA adapter compatibility.
+    The ibm-granite/rag-intrinsics-lib repository only has adapters for
+    Granite 3.3 models. Granite 4 adapters are not yet available.
+    """
     backend = LocalHFBackend(
         model_id="ibm-granite/granite-3.3-8b-instruct",
         formatter=TemplateFormatter(model_id="ibm-granite/granite-4.0-tiny-preview"),
diff --git a/test/backends/test_litellm_ollama.py b/test/backends/test_litellm_ollama.py
index 7999528a..ece9c890 100644
--- a/test/backends/test_litellm_ollama.py
+++ b/test/backends/test_litellm_ollama.py
@@ -14,7 +14,7 @@
 from mellea.stdlib.context import SimpleContext
 from mellea.stdlib.sampling import RejectionSamplingStrategy
 
-_MODEL_ID = f"ollama_chat/{model_ids.IBM_GRANITE_4_MICRO_3B.ollama_name}"
+_MODEL_ID = f"ollama_chat/{model_ids.IBM_GRANITE_4_HYBRID_MICRO.ollama_name}"
 
 
 @pytest.fixture(scope="function")
diff --git a/test/backends/test_openai_ollama.py b/test/backends/test_openai_ollama.py
index d95648b0..95989729 100644
--- a/test/backends/test_openai_ollama.py
+++ b/test/backends/test_openai_ollama.py
@@ -12,7 +12,7 @@
 
 from mellea import MelleaSession
 from mellea.backends import ModelOption
-from mellea.backends.model_ids import IBM_GRANITE_4_MICRO_3B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.backends.openai import OpenAIBackend
 from mellea.core import CBlock, ModelOutputThunk
 from mellea.formatters import TemplateFormatter
@@ -23,8 +23,8 @@
 def backend(gh_run: int):
     """Shared OpenAI backend configured for Ollama."""
     return OpenAIBackend(
-        model_id=IBM_GRANITE_4_MICRO_3B.ollama_name,  # type: ignore
-        formatter=TemplateFormatter(model_id=IBM_GRANITE_4_MICRO_3B.hf_model_name),  # type: ignore
+        model_id=IBM_GRANITE_4_HYBRID_MICRO.ollama_name,  # type: ignore
+        formatter=TemplateFormatter(model_id=IBM_GRANITE_4_HYBRID_MICRO.hf_model_name),  # type: ignore
         base_url=f"http://{os.environ.get('OLLAMA_HOST', 'localhost:11434')}/v1",
         api_key="ollama",
     )
diff --git a/test/backends/test_openai_vllm/test_openai_vllm.py b/test/backends/test_openai_vllm/test_openai_vllm.py
index 2029dfe5..ad55fe39 100644
--- a/test/backends/test_openai_vllm/test_openai_vllm.py
+++ b/test/backends/test_openai_vllm/test_openai_vllm.py
@@ -133,7 +133,7 @@ class Answer(pydantic.BaseModel):
 class TestOpenAIALoraStuff:
     backend = OpenAIBackend(
         model_id="ibm-granite/granite-3.3-8b-instruct",
-        formatter=TemplateFormatter(model_id="ibm-granite/granite-4.0-tiny-preview"),
+        formatter=TemplateFormatter(model_id="ibm-granite/granite-3.3-8b-instruct"),
         base_url="http://localhost:8000/v1",
         api_key="EMPTY",
     )
diff --git a/test/backends/test_vision_ollama.py b/test/backends/test_vision_ollama.py
index e157d0e5..21782b1d 100644
--- a/test/backends/test_vision_ollama.py
+++ b/test/backends/test_vision_ollama.py
@@ -9,7 +9,7 @@
 pytestmark = [pytest.mark.ollama, pytest.mark.llm]
 
 from mellea import MelleaSession, start_session
-from mellea.backends import ModelOption
+from mellea.backends import ModelOption, model_ids
 from mellea.core import ImageBlock, ModelOutputThunk
 from mellea.stdlib.components import Instruction, Message
 
diff --git a/test/backends/test_vision_openai.py b/test/backends/test_vision_openai.py
index bfcfd681..599d6cbf 100644
--- a/test/backends/test_vision_openai.py
+++ b/test/backends/test_vision_openai.py
@@ -11,7 +11,7 @@
 
 from mellea import MelleaSession, start_session
 from mellea.backends import ModelOption
-from mellea.backends.model_ids import IBM_GRANITE_4_MICRO_3B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.core import ImageBlock, ModelOutputThunk
 from mellea.stdlib.components import Instruction, Message
 
@@ -21,7 +21,7 @@ def m_session(gh_run):
     if gh_run == 1:
         m = start_session(
             "openai",
-            model_id=IBM_GRANITE_4_MICRO_3B.ollama_name,  # type: ignore
+            model_id=IBM_GRANITE_4_HYBRID_MICRO.ollama_name,  # type: ignore
             base_url=f"http://{os.environ.get('OLLAMA_HOST', 'localhost:11434')}/v1",
             api_key="ollama",
             model_options={ModelOption.MAX_NEW_TOKENS: 5},
diff --git a/test/backends/test_watsonx.py b/test/backends/test_watsonx.py
index 6902e675..306bd5d5 100644
--- a/test/backends/test_watsonx.py
+++ b/test/backends/test_watsonx.py
@@ -18,7 +18,7 @@
 ]
 
 from mellea import MelleaSession
-from mellea.backends import ModelOption
+from mellea.backends import ModelOption, model_ids
 from mellea.backends.watsonx import WatsonxAIBackend
 from mellea.core import CBlock, ModelOutputThunk
 from mellea.formatters import TemplateFormatter
@@ -32,8 +32,8 @@ def backend():
         pytest.skip("Skipping watsonx tests.")
     else:
         return WatsonxAIBackend(
-            model_id="ibm/granite-3-3-8b-instruct",
-            formatter=TemplateFormatter(model_id="ibm-granite/granite-3.3-8b-instruct"),
+            model_id=model_ids.IBM_GRANITE_4_HYBRID_SMALL,
+            formatter=TemplateFormatter(model_id=model_ids.IBM_GRANITE_4_HYBRID_SMALL),
         )
 
 
@@ -223,6 +223,21 @@ async def get_client_async():
     assert len(backend._client_cache.cache.values()) == 2
 
 
+def test_default_model():
+    """Verify WatsonxAIBackend uses correct default model."""
+    if int(os.environ.get("CICD", 0)) == 1:
+        pytest.skip("Skipping watsonx tests.")
+
+    # Create backend without specifying model_id
+    default_backend = WatsonxAIBackend()
+
+    # Verify it uses IBM_GRANITE_4_HYBRID_SMALL as default
+    assert default_backend._model_id == model_ids.IBM_GRANITE_4_HYBRID_SMALL, (
+        f"Expected default model to be IBM_GRANITE_4_HYBRID_SMALL, "
+        f"but got {default_backend._model_id}"
+    )
+
+
 if __name__ == "__main__":
     import pytest
 
diff --git a/test/core/test_component_typing.py b/test/core/test_component_typing.py
index a20de283..19ca2d23 100644
--- a/test/core/test_component_typing.py
+++ b/test/core/test_component_typing.py
@@ -6,7 +6,7 @@
 
 import mellea.stdlib.functional as mfuncs
 from mellea import MelleaSession, start_session
-from mellea.backends.model_ids import IBM_GRANITE_4_MICRO_3B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.backends.ollama import OllamaModelBackend
 from mellea.core import (
     CBlock,
@@ -64,10 +64,10 @@ def backend(gh_run: int):
     """Shared backend."""
     if gh_run == 1:
         return OllamaModelBackend(
-            model_id=IBM_GRANITE_4_MICRO_3B.ollama_name  # type: ignore
+            model_id=IBM_GRANITE_4_HYBRID_MICRO.ollama_name  # type: ignore
         )
     else:
-        return OllamaModelBackend(model_id="granite3.3:8b")
+        return OllamaModelBackend(model_id=IBM_GRANITE_4_HYBRID_MICRO.ollama_name)  # type: ignore
 
 
 @pytest.fixture(scope="module")
@@ -117,11 +117,10 @@ def test_incorrect_type_override():
 
 
 # Marking as qualitative for now since there's so much generation required for this.
-# Uses granite3.3:8b (8B, heavy) in local mode
+# Uses granite4:micro-h (3B hybrid, lightweight) in local mode
 @pytest.mark.qualitative
 @pytest.mark.ollama
 @pytest.mark.requires_gpu
-@pytest.mark.requires_heavy_ram
 @pytest.mark.llm
 async def test_generating(session):
     m = session
diff --git a/test/formatters/test_template_formatter.py b/test/formatters/test_template_formatter.py
index 5d851053..5dd995b8 100644
--- a/test/formatters/test_template_formatter.py
+++ b/test/formatters/test_template_formatter.py
@@ -5,7 +5,7 @@
 
 import pytest
 
-from mellea.backends.model_ids import IBM_GRANITE_3_2_8B, ModelIdentifier
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO, ModelIdentifier
 from mellea.core import CBlock, Component, ModelOutputThunk, TemplateRepresentation
 from mellea.formatters import TemplateFormatter
 from mellea.stdlib.components import Instruction, Message, MObject
@@ -173,7 +173,7 @@ def _parse(self, computed: ModelOutputThunk) -> str:
 
 
 def test_load_with_model_id(instr: Instruction):
-    tf = TemplateFormatter(IBM_GRANITE_3_2_8B)
+    tf = TemplateFormatter(IBM_GRANITE_4_HYBRID_MICRO)
     tmpl = tf._load_template(instr.format_for_llm())
     assert tmpl.name is not None
     assert "granite" in tmpl.name, (
diff --git a/test/stdlib/components/test_genslot.py b/test/stdlib/components/test_genslot.py
index d3814ae4..9f63d4d7 100644
--- a/test/stdlib/components/test_genslot.py
+++ b/test/stdlib/components/test_genslot.py
@@ -4,7 +4,7 @@
 import pytest
 
 from mellea import MelleaSession, generative, start_session
-from mellea.backends.model_ids import IBM_GRANITE_4_MICRO_3B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.backends.ollama import OllamaModelBackend
 from mellea.core import Requirement
 from mellea.stdlib.components.genslot import (
@@ -17,13 +17,8 @@
 from mellea.stdlib.requirements import simple_validate
 from mellea.stdlib.sampling import RejectionSamplingStrategy
 
-# Module-level markers: Uses granite3.3:8b (8B, heavy) in local mode
-pytestmark = [
-    pytest.mark.ollama,
-    pytest.mark.requires_gpu,
-    pytest.mark.requires_heavy_ram,
-    pytest.mark.llm,
-]
+# Module-level markers: Uses granite4:micro-h (3B hybrid, lightweight) in local mode
+pytestmark = [pytest.mark.ollama, pytest.mark.requires_gpu, pytest.mark.llm]
 
 
 @pytest.fixture(scope="module")
@@ -31,10 +26,10 @@ def backend(gh_run: int):
     """Shared backend."""
     if gh_run == 1:
         return OllamaModelBackend(
-            model_id=IBM_GRANITE_4_MICRO_3B.ollama_name  # type: ignore
+            model_id=IBM_GRANITE_4_HYBRID_MICRO.ollama_name  # type: ignore
         )
     else:
-        return OllamaModelBackend(model_id="granite3.3:8b")
+        return OllamaModelBackend(model_id=IBM_GRANITE_4_HYBRID_MICRO.ollama_name)  # type: ignore
 
 
 @generative
diff --git a/test/stdlib/test_spans.py b/test/stdlib/test_spans.py
index 311ec3c1..06c9f8a2 100644
--- a/test/stdlib/test_spans.py
+++ b/test/stdlib/test_spans.py
@@ -2,18 +2,13 @@
 
 from mellea.backends import ModelOption
 from mellea.backends.huggingface import LocalHFBackend
-from mellea.backends.model_ids import IBM_GRANITE_3_3_8B
+from mellea.backends.model_ids import IBM_GRANITE_4_HYBRID_MICRO
 from mellea.core import CBlock
 from mellea.stdlib.components import SimpleComponent
 from mellea.stdlib.session import MelleaSession, start_session
 
-# Module-level markers for all tests using granite-3.3-8b (8B model)
-pytestmark = [
-    pytest.mark.huggingface,
-    pytest.mark.requires_gpu,
-    pytest.mark.requires_heavy_ram,
-    pytest.mark.llm,
-]
+# Module-level markers for all tests using Granite 4 hybrid micro (3B model)
+pytestmark = [pytest.mark.huggingface, pytest.mark.requires_gpu, pytest.mark.llm]
 
 
 # We edit the context type in the async tests below. Don't change the scope here.
@@ -21,7 +16,7 @@
 def m_session(gh_run):
     m = start_session(
         "hf",
-        model_id=IBM_GRANITE_3_3_8B,
+        model_id=IBM_GRANITE_4_HYBRID_MICRO,
         model_options={ModelOption.MAX_NEW_TOKENS: 64},
     )
     yield m