Default to ollama:qwen3.8

Replaces gpt-oss on ModelConfig, qa.model and processing.title_model, and
ministral-3 on the picture-description model. qa.model.vision follows the
model and is now true.

enable_thinking was gated on the gpt-oss name, so it did nothing for
qwen3.8. With title_model's max_tokens of 100 the reasoning consumed the
whole budget and title generation returned an empty string. The mapping
now applies to any ollama model via reasoning_effort(): false sends
"none", true sends "high". Measured on qwen3.8:27b-mlx, "low" does not
disable thinking and "none" does; gpt-oss is the inverse, its template
has no "none" level, so it keeps "low".

Picture description bypasses get_model -- docling posts the request
itself from a params dict -- so the flag was inert on that path too.
vlm_api_params() carries reasoning_effort into both converters' request
bodies. At max_tokens 200 the description survived either way, but the
switch cut completion tokens from 141 to 45.

test_search_tool_skips_binary_content_when_qa_model_is_text_only asserted
the vision default rather than setting it; it now configures vision=False
itself.

docs/benchmarks.md keeps ministral-3: those are recorded measurements.

CHANGELOG.md

@@ -2,6 +2,21 @@
 
 ## [Unreleased]
 
+### Changed
+
+- Default models are `ollama:qwen3.8`: `ModelConfig`, `qa.model`,
+  `processing.title_model` (was `ollama:gpt-oss`) and
+  `processing.conversion_options.picture_description.model` (was
+  `ollama:ministral-3`). Run `ollama pull qwen3.8`.
+- `qa.model.vision` defaults to `true`, matching `qwen3.8`. Set it `false` when
+  pointing `qa.model` at a text-only model.
+- `enable_thinking` on `provider: ollama` maps to `reasoning_effort` for every
+  model, not only `gpt-oss`: `false` sends `none`, `true` sends `high`.
+  `gpt-oss` keeps `low` for `false`.
+- `processing.conversion_options.picture_description.model` defaults to
+  `enable_thinking: false`, and the field now reaches the VLM: docling's
+  picture-description request carries `reasoning_effort` in `params`.
+
 ## [0.82.1] - 2026-09-03
 
 ### Fixed

docs/configuration/index.md

@@ -61,7 +61,7 @@ embeddings:
 qa:
   model:
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
     enable_thinking: true
 ```
 
@@ -105,7 +105,7 @@ reranking:
 qa:
   model:
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
     enable_thinking: true
     temperature: 0.3
   max_searches: 5
@@ -135,7 +135,7 @@ processing:
   auto_title: false              # Auto-generate titles on ingestion
   title_model:
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
     enable_thinking: false
     temperature: 0.3
     max_tokens: 100

docs/configuration/processing.md

@@ -30,7 +30,7 @@ processing:
   auto_title: false                          # Auto-generate titles on ingestion
   title_model:                               # LLM for title generation (fallback)
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
     enable_thinking: false
 
   # Conversion options (works with both local and remote converters)
@@ -54,7 +54,7 @@ processing:
     picture_description:
       model:
         provider: ollama
-        name: ministral-3
+        name: qwen3.8
   pictures: image                            # none | description | image
 ```
 
@@ -270,7 +270,7 @@ processing:
     picture_description:          # only consulted when pictures == "description"
       model:
         provider: ollama          # any OpenAI-compatible /v1/chat/completions provider
-        name: ministral-3
+        name: qwen3.8
       timeout: 90
       max_tokens: 200
 ```
@@ -294,7 +294,7 @@ Three independent settings drive ingest, retrieval, and QA:
 |---|---|---|
 | `processing.pictures` | Generate and/or describe pictures at ingest? | `none` / `description` / `image` (default) |
 | `embeddings.model.multimodal` | Can the embedder index image content? | `false` (default, text-only) / `true` (supported on `vllm`, `voyageai`, `cohere`) |
-| `qa.model.vision` | Can the QA model interpret images? | `false` (default) / `true` |
+| `qa.model.vision` | Can the QA model interpret images? | `false` / `true` (default) |
 
 The Embedder column below is driven by `embeddings.model.multimodal`, not the provider name — a vision-capable model under a text-only configuration still indexes no images, and an image-only document then produces zero chunks. See [Multimodal embedders](providers.md#multimodal-embedders).
 
@@ -313,7 +313,7 @@ The Embedder column below is driven by `embeddings.model.multimodal`, not the pr
 - `qa.model.vision: false` — text chunks only (descriptions, when present, answer figure questions in prose).
 - `qa.model.vision: true` — text chunks + raw picture bytes via `BinaryContent`. The model reads figures directly. Requires `pictures != none` so the bytes exist.
 
-`qa.model.vision` is independent of ingestion. Flipping it never requires reingesting. Setting `vision: true` against a text-only model causes silent acceptance and confabulation on Ollama and a 400 on OpenAI. Default `false` is the safe choice.
+`qa.model.vision` is independent of ingestion. Flipping it never requires reingesting. It declares what the model can read: the default `qwen3.8` is vision-capable, so the default is `true`. Set it `false` when pointing `qa.model` at a text-only model, where `true` causes silent acceptance and confabulation on Ollama and a 400 on OpenAI.
 
 **Recommended combinations:**
 
@@ -368,7 +368,7 @@ processing:
   auto_title: true
   title_model:
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
     enable_thinking: false
 ```
 

docs/configuration/providers.md

@@ -15,7 +15,7 @@ Configure model behavior for the `qa` and `analysis` capabilities. These setting
 qa:
   model:
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
     temperature: 0.3
     max_tokens: 500
 ```
@@ -79,7 +79,7 @@ See the [Pydantic AI thinking documentation](https://ai.pydantic.dev/thinking/)
 - **Google**: Gemini models with thinking support
 - **Groq**: Models with reasoning capabilities
 - **Bedrock**: Claude, Qwen, and `gpt-oss` models. Bedrock Converse does not serve the proprietary OpenAI models, so configuring one raises an error. Reach those through `provider: bedrock-mantle`.
-- **Ollama**: Models supporting reasoning (gpt-oss, etc.)
+- **Ollama**: Any model with a thinking capability. `enable_thinking` maps to `reasoning_effort`: `false` sends `none` (`low` for `gpt-oss`, whose template has no `none` level), `true` sends `high`.
 - **vLLM**: Models with a pydantic-ai reasoning profile (gpt-oss). Qwen3, Gemma, and similar templates ignore the OpenAI `reasoning_effort` that `enable_thinking` translates to — use [`extra_body`](#raw-provider-pass-through) to drive them.
 - **LM Studio**: Models supporting reasoning (gpt-oss, etc.)
 
@@ -311,7 +311,7 @@ Configure which LLM provider to use for question answering. Any provider and mod
 qa:
   model:
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
 ```
 
 The Ollama base URL can be configured via the `OLLAMA_BASE_URL` environment variable, config file, or defaults to `http://localhost:11434`:

docs/configuration/qa.md

@@ -26,10 +26,10 @@ Configure the RAG capability (used by `client.ask`, `haiku-rag ask`, and the MCP
 qa:
   model:
     provider: ollama
-    name: gpt-oss
+    name: qwen3.8
     enable_thinking: true
     temperature: 0.3          # Default: 0.3
-    vision: false             # Set true for vision-capable models
+    vision: true              # Set false for text-only models
   max_searches: 5       # Maximum search units per question
 ```
 

docs/tutorial.md

@@ -12,7 +12,7 @@ You also need [Ollama](https://ollama.com/) for the default embedding and answer
 
 ```bash
 ollama pull qwen3-embedding:4b
-ollama pull gpt-oss
+ollama pull qwen3.8
 ```
 
 !!! note "Prefer OpenAI?"

evaluations/README.md

@@ -49,7 +49,7 @@ datasets and judge:
 
 ```bash
 evaluations run hotpotqa --target rag-capability
-evaluations run hotpotqa --target analysis-capability --capability-model ollama:gpt-oss
+evaluations run hotpotqa --target analysis-capability --capability-model ollama:qwen3.8
 ```
 
 `--capability-model "provider:name"` overrides the capability model independently from

haiku_rag_slim/haiku/rag/config/models.py

@@ -40,7 +40,7 @@ class ModelConfig(ConfigModel):
     """
 
     provider: str = "ollama"
-    name: str = "gpt-oss"
+    name: str = "qwen3.8"
     base_url: str | None = None
     api_key: str | None = None
 
@@ -165,9 +165,10 @@ class QAConfig(ConfigModel):
     model: ModelConfig = Field(
         default_factory=lambda: ModelConfig(
             provider="ollama",
-            name="gpt-oss",
+            name="qwen3.8",
             enable_thinking=True,
             temperature=0.3,
+            vision=True,
         )
     )
     max_searches: int = Field(default=5, ge=0)
@@ -216,7 +217,8 @@ class PictureDescriptionConfig(ConfigModel):
     model: ModelConfig = Field(
         default_factory=lambda: ModelConfig(
             provider="ollama",
-            name="ministral-3",
+            name="qwen3.8",
+            enable_thinking=False,
             temperature=0.0,
         )
     )
@@ -303,7 +305,7 @@ class ProcessingConfig(ConfigModel):
     title_model: ModelConfig = Field(
         default_factory=lambda: ModelConfig(
             provider="ollama",
-            name="gpt-oss",
+            name="qwen3.8",
             enable_thinking=False,
             temperature=0.3,
             max_tokens=100,

haiku_rag_slim/haiku/rag/converters/base.py

@@ -42,6 +42,20 @@ def vlm_api_headers(model: "ModelConfig") -> dict[str, str]:
     return {}
 
 
+def vlm_api_params(model: "ModelConfig", max_tokens: int) -> dict[str, object]:
+    """Request body fields docling posts alongside the picture."""
+    from haiku.rag.utils import reasoning_effort
+
+    params: dict[str, object] = {
+        "model": model.name,
+        "max_completion_tokens": max_tokens,
+    }
+    effort = reasoning_effort(model)
+    if effort is not None:
+        params["reasoning_effort"] = effort
+    return params
+
+
 class DocumentConverter(ABC):
     """Abstract base class for document converters.
 

haiku_rag_slim/haiku/rag/converters/docling_local.py

@@ -12,6 +12,7 @@ from haiku.rag.config import AppConfig
 from haiku.rag.converters.base import (
     DocumentConverter,
     vlm_api_headers,
+    vlm_api_params,
     vlm_api_url,
 )
 from haiku.rag.converters.text_utils import TextFileHandler, docling_safe_name
@@ -153,10 +154,7 @@ class DoclingLocalConverter(DocumentConverter):
             pipeline_options.picture_description_options = PictureDescriptionApiOptions(
                 url=AnyUrl(vlm_api_url(self.config, pic_desc.model)),
                 headers=vlm_api_headers(pic_desc.model),
-                params=dict(
-                    model=pic_desc.model.name,
-                    max_completion_tokens=pic_desc.max_tokens,
-                ),
+                params=vlm_api_params(pic_desc.model, pic_desc.max_tokens),
                 prompt=self.config.prompts.picture_description,
                 timeout=pic_desc.timeout,
             )

haiku_rag_slim/haiku/rag/converters/docling_serve.py

@@ -9,6 +9,7 @@ from haiku.rag.config import AppConfig
 from haiku.rag.converters.base import (
     DocumentConverter,
     vlm_api_headers,
+    vlm_api_params,
     vlm_api_url,
 )
 from haiku.rag.converters.text_utils import TextFileHandler, docling_safe_name
@@ -110,10 +111,7 @@ class DoclingServeConverter(DocumentConverter):
             picture_description_api = {
                 "url": vlm_api_url(self.config, pic_desc.model),
                 "headers": vlm_api_headers(pic_desc.model),
-                "params": {
-                    "model": pic_desc.model.name,
-                    "max_completion_tokens": pic_desc.max_tokens,
-                },
+                "params": vlm_api_params(pic_desc.model, pic_desc.max_tokens),
                 "prompt": prompt,
                 "timeout": pic_desc.timeout,
             }

haiku_rag_slim/haiku/rag/utils.py

@@ -4,7 +4,7 @@ import sys
 from collections.abc import Awaitable
 from importlib import metadata
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, NoReturn, cast
+from typing import TYPE_CHECKING, Any, Literal, NoReturn, cast
 
 from packaging.version import Version, parse
 
@@ -41,7 +41,7 @@ def parse_model_option(value: str) -> "ModelConfig":
     parts = value.split(":", 1)
     if len(parts) != 2 or not parts[0] or not parts[1]:
         raise ValueError(
-            f"Invalid model format '{value}'. Expected 'provider:name' (e.g. 'ollama:gpt-oss')."
+            f"Invalid model format '{value}'. Expected 'provider:name' (e.g. 'ollama:qwen3.8')."
         )
     return ModelConfig(provider=parts[0], name=parts[1])
 
@@ -182,6 +182,20 @@ _OPENAI_COMPAT_PROFILE: "OpenAIModelProfile" = {
 }
 
 
+def reasoning_effort(
+    model_config: "ModelConfig",
+) -> Literal["none", "low", "high"] | None:
+    """OpenAI `reasoning_effort` for a model config, or None when unset.
+
+    "low" is gpt-oss's floor; its template rejects "none".
+    """
+    if model_config.enable_thinking is None:
+        return None
+    if model_config.enable_thinking:
+        return "high"
+    return "low" if model_config.name == "gpt-oss" else "none"
+
+
 def get_model(
     model_config: "ModelConfig",
     app_config: "AppConfig | None" = None,
@@ -213,12 +227,9 @@ def get_model(
     if provider == "ollama":
         model_settings = None
 
-        # Apply thinking control for gpt-oss
-        if model == "gpt-oss" and model_config.enable_thinking is not None:
-            if model_config.enable_thinking is False:
-                model_settings = OpenAIChatModelSettings(openai_reasoning_effort="low")
-            else:
-                model_settings = OpenAIChatModelSettings(openai_reasoning_effort="high")
+        effort = reasoning_effort(model_config)
+        if effort is not None:
+            model_settings = OpenAIChatModelSettings(openai_reasoning_effort=effort)
 
         model_settings = apply_common_settings(
             model_settings, model_config, map_thinking=False

tests/conftest.py

@@ -102,7 +102,7 @@ def temp_yaml_config(tmp_path, monkeypatch):
                 "vector_dim": 2560,
             }
         },
-        "qa": {"model": {"provider": "ollama", "name": "gpt-oss"}},
+        "qa": {"model": {"provider": "ollama", "name": "qwen3.8"}},
     }
 
     with open(config_file, "w") as f:

tests/test_converters.py

@@ -15,7 +15,11 @@ from docling_core.types.doc.document import DoclingDocument
 from haiku.rag.config import AppConfig
 from haiku.rag.config.models import ModelConfig
 from haiku.rag.converters import docling_local, get_converter
-from haiku.rag.converters.base import vlm_api_headers, vlm_api_url
+from haiku.rag.converters.base import (
+    vlm_api_headers,
+    vlm_api_params,
+    vlm_api_url,
+)
 from haiku.rag.converters.docling_local import DoclingLocalConverter
 from haiku.rag.converters.docling_serve import DoclingServeConverter
 from haiku.rag.converters.text_utils import TextFileHandler, docling_safe_name
@@ -50,6 +54,32 @@ class TestVlmApiUrl:
             vlm_api_url(AppConfig(), ModelConfig(provider="unsupported", name="test"))
 
 
+class TestVlmApiParams:
+    """Request body docling posts alongside the picture."""
+
+    def test_thinking_unset_sends_no_effort(self):
+        params = vlm_api_params(ModelConfig(provider="ollama", name="qwen3.8"), 200)
+        assert params == {"model": "qwen3.8", "max_completion_tokens": 200}
+
+    def test_thinking_off_sends_none(self):
+        params = vlm_api_params(
+            ModelConfig(provider="ollama", name="qwen3.8", enable_thinking=False), 200
+        )
+        assert params["reasoning_effort"] == "none"
+
+    def test_thinking_off_sends_gpt_oss_floor(self):
+        params = vlm_api_params(
+            ModelConfig(provider="ollama", name="gpt-oss", enable_thinking=False), 200
+        )
+        assert params["reasoning_effort"] == "low"
+
+    def test_thinking_on_sends_high(self):
+        params = vlm_api_params(
+            ModelConfig(provider="ollama", name="qwen3.8", enable_thinking=True), 200
+        )
+        assert params["reasoning_effort"] == "high"
+
+
 class TestVlmApiHeaders:
     """Auth headers for picture-description VLM models."""
 
@@ -989,7 +1019,7 @@ class TestDoclingLocalConverter:
         pic_desc = config.processing.conversion_options.picture_description
         assert config.processing.pictures == "image"
         assert pic_desc.model.provider == "ollama"
-        assert pic_desc.model.name == "ministral-3"
+        assert pic_desc.model.name == "qwen3.8"
         assert pic_desc.timeout == 90
         assert pic_desc.max_tokens == 200
         # Default prompt is in PromptsConfig

tests/test_doctor.py

@@ -163,7 +163,7 @@ def _stub_provider_probe(monkeypatch):
             {
                 "models": [
                     {"name": "test"},
-                    {"name": "gpt-oss:latest"},
+                    {"name": "qwen3.8:latest"},
                     {"name": "qwen3-embedding:4b"},
                 ]
             },
@@ -751,14 +751,11 @@ def test_api_key_not_required_when_config_supplies_it():
 
 
 def test_active_models_includes_picture_description_when_enabled():
-    config = AppConfig(processing=ProcessingConfig(pictures="description"))
-    names = [model.name for model in _active_models(config)]
-    assert "ministral-3" in names
-
-
-def test_active_models_excludes_picture_description_by_default():
-    names = [model.name for model in _active_models(AppConfig())]
-    assert "ministral-3" not in names
+    base = _active_models(AppConfig())
+    with_pictures = _active_models(
+        AppConfig(processing=ProcessingConfig(pictures="description"))
+    )
+    assert len(with_pictures) == len(base) + 1
 
 
 def test_active_models_includes_title_model_when_auto_title():
@@ -846,7 +843,7 @@ def test_provider_targets_default_groups_ollama_models():
     assert len(targets) == 1
     entry = next(iter(targets.values()))
     assert entry["kind"] == "ollama"
-    assert {"qwen3-embedding:4b", "gpt-oss"} <= entry["models"]
+    assert {"qwen3-embedding:4b", "qwen3.8"} <= entry["models"]
 
 
 def test_provider_targets_includes_docling_serve():
@@ -922,7 +919,7 @@ async def test_provider_check_ok_when_models_present(monkeypatch):
                 {
                     "models": [
                         {"name": "qwen3-embedding:4b"},
-                        {"name": "gpt-oss:latest"},
+                        {"name": "qwen3.8:latest"},
                     ]
                 },
             )
@@ -960,7 +957,7 @@ async def test_provider_check_fails_when_unreachable(monkeypatch):
 async def test_provider_check_reports_local_provider(monkeypatch):
     monkeypatch.setattr(
         "haiku.rag.doctor._probe_endpoint",
-        _fake_probe((True, None, {"models": [{"name": "gpt-oss:latest"}]})),
+        _fake_probe((True, None, {"models": [{"name": "qwen3.8:latest"}]})),
     )
     config = AppConfig(
         embeddings=EmbeddingsConfig(
@@ -981,7 +978,7 @@ async def test_run_doctor_includes_provider_results(temp_db_path, monkeypatch):
     monkeypatch.setattr(
         "haiku.rag.doctor._probe_endpoint",
         _fake_probe(
-            (True, None, {"models": [{"name": "test"}, {"name": "gpt-oss:latest"}]})
+            (True, None, {"models": [{"name": "test"}, {"name": "qwen3.8:latest"}]})
         ),
     )
     report = await run_doctor(_config(), temp_db_path, {})

tests/test_download_models.py

@@ -77,8 +77,8 @@ async def test_download_models_ollama_pulls_models(mock_to_thread):
         async for progress in download_models(get_config()):
             events.append(progress)
 
-    # Default config has embeddings=qwen3-embedding:4b, qa=gpt-oss
-    ollama_models = {"gpt-oss", "qwen3-embedding:4b"}
+    # Default config has embeddings=qwen3-embedding:4b, qa=qwen3.8
+    ollama_models = {"qwen3.8", "qwen3-embedding:4b"}
     ollama_events = [e for e in events if e.model in ollama_models]
     pulling_events = [e for e in ollama_events if e.status == "pulling"]
     done_events = [e for e in ollama_events if e.status == "done"]
@@ -108,7 +108,7 @@ async def test_download_models_no_ollama_models(mock_to_thread):
 
     models = {e.model for e in events}
     assert "qwen3-embedding:4b" not in models
-    assert "gpt-oss" not in models
+    assert "qwen3.8" not in models
 
 
 @pytest.mark.parametrize(

tests/test_picture_in_context.py

@@ -837,7 +837,7 @@ async def test_ingest_emits_picture_chunks_with_multimodal_embedder(
 @pytest.mark.asyncio
 async def test_search_tool_skips_binary_content_when_qa_model_is_text_only():
     """The agent search tool must NOT attach picture bytes when the QA model
-    is text-only (``qa.model.vision = False``, the default). Sending image
+    is text-only (``qa.model.vision = False``). Sending image
     parts to a text-only model would cause it to hallucinate confidently —
     Ollama silently accepts the bytes and the model guesses."""
 
@@ -856,8 +856,7 @@ async def test_search_tool_skips_binary_content_when_qa_model_is_text_only():
     fake_client.expand_context = AsyncMock(return_value=[picture_result])
 
     config = AppConfig()
-    # vision defaults to False; assert anyway so the test reads explicitly.
-    assert config.qa.model.vision is False
+    config.qa.model.vision = False
     toolset = create_search_toolset(config, expand_context=False)
     func = toolset.tools["search"].function
 

tests/test_utils.py

@@ -154,6 +154,14 @@ Emoji test: 🚀 ✅ 📝"""
             {"provider": "ollama", "name": "gpt-oss", "enable_thinking": True},
             {"openai_reasoning_effort": "high"},
         ),
+        (
+            {"provider": "ollama", "name": "qwen3.8", "enable_thinking": False},
+            {"openai_reasoning_effort": "none"},
+        ),
+        (
+            {"provider": "ollama", "name": "qwen3.8", "enable_thinking": True},
+            {"openai_reasoning_effort": "high"},
+        ),
         (
             {
                 "provider": "ollama",
@@ -188,6 +196,8 @@ Emoji test: 🚀 ✅ 📝"""
         "ollama",
         "ollama_thinking_off",
         "ollama_thinking_on",
+        "ollama_other_thinking_off",
+        "ollama_other_thinking_on",
         "ollama_with_settings",
         "openai",
         "openai_reasoning_thinking_on",