getcellm · kaspermarstal · Apr 8, 2026 · Apr 8, 2026 · Apr 8, 2026
diff --git a/README.md b/README.md
@@ -47,7 +47,7 @@ Just remember that the models do make mistakes at times. They might misunderstan
 
 3. Open Excel, choose a provider from the drop-down menu in the Cellm tab, and plug in your API key.
 
-You can also use local models, e.g., via [Ollama](https://ollama.com/). Download and install [Ollama](https://ollama.com/), open Windows Terminal (open start menu, type `Windows Terminal`, and click `OK`), type `ollama pull gemma3:4b`, and wait for the download to finish. Open Excel, choose the Ollama provider from the drop-down menu in the Cellm tab, and you are good to go.
+You can also use local models, e.g., via [Ollama](https://ollama.com/). Download and install [Ollama](https://ollama.com/) and open Excel. Choose the Ollama provider from the drop-down menu in the Cellm tab and select a model. Cellm will prompt you to download it automatically. Alternatively, open Windows Terminal (open start menu, type `Windows Terminal`, and click `OK`), type `ollama pull gemma4:e4b`, and wait for the download to finish.
 
 ## Pricing
 - **Free tier:** Use local models or your own API keys 
@@ -56,7 +56,7 @@ You can also use local models, e.g., via [Ollama](https://ollama.com/). Download
 
 ## Basic usage
 
-Select a cell and type `=PROMPT("What model are you and who made you?")`. For Gemma 3 4B, it will tell you that it's called "Gemma" and made by Google DeepMind.
+Select a cell and type `=PROMPT("What model are you and who made you?")`. For Gemma 4 E4B, it will tell you that it's called "Gemma 4" and made by Google DeepMind.
 
 You can also use cell references. For example, copy a news article into cell A1 and type in cell B1: `=PROMPT("Extract all person names mentioned in the text", A1)`. You can reference many cells using standard Excel notation, e.g. `=PROMPT("Extract all person names in the cells", A1:F10)` or reference multiple separate ranges, e.g. `=PROMPT("Compare these datasets", A1:B10, D1:E10)`
 

diff --git a/docs/api-reference/functions/prompt-model.mdx b/docs/api-reference/functions/prompt-model.mdx
@@ -7,7 +7,7 @@ Allows you to call a model from a cell formula and specify the model as the firs
 ## Arguments
 
 <ParamField body="providerAndModel" type="string" required>
-  A string on the form "provider/model" (e.g., "openai/gpt-4o-mini").
+  A string on the form "provider/model" (e.g., "openai/gpt-5.4-mini").
 
   The default model is determined by your configuration settings.
 </ParamField>
@@ -44,39 +44,39 @@ Allows you to call a model from a cell formula and specify the model as the firs
 <RequestExample>
 
 ```excel Text Instructions
-=PROMPTMODEL("openai/gpt-4o-mini", "Extract keywords")
+=PROMPTMODEL("openai/gpt-5.4-mini", "Extract keywords")
 ```
 
 ```excel Cell Instructions
-=PROMPTMODEL("openai/gpt-4o-mini", A1:D10)
+=PROMPTMODEL("openai/gpt-5.4-mini", A1:D10)
 ```
 
 ```excel With Context
-=PROMPTMODEL("openai/gpt-4o-mini", "Extract keywords", A1:D10)
+=PROMPTMODEL("openai/gpt-5.4-mini", "Extract keywords", A1:D10)
 ```
 
 ```excel Multiple Cell Ranges
-=PROMPTMODEL("openai/gpt-4o-mini", "Compare these datasets", A1:B10, D1:E10)
+=PROMPTMODEL("openai/gpt-5.4-mini", "Compare these datasets", A1:B10, D1:E10)
 ```
 
 ```excel Mixed Cell References
-=PROMPTMODEL("openai/gpt-4o-mini", "Analyze all data", A1, B2:C5, D6)
+=PROMPTMODEL("openai/gpt-5.4-mini", "Analyze all data", A1, B2:C5, D6)
 ```
 
 </RequestExample>
 
 <ResponseExample>
 
 ```excel TOROW
-=PROMPTMODEL.TOROW("openai/gpt-4o-mini", "Extract keywords", A1:D10)
+=PROMPTMODEL.TOROW("openai/gpt-5.4-mini", "Extract keywords", A1:D10)
 ```
 
 ```excel TOCOLUMN
-=PROMPTMODEL.TOCOLUMN("openai/gpt-4o-mini", "Extract keywords", A1:D10)
+=PROMPTMODEL.TOCOLUMN("openai/gpt-5.4-mini", "Extract keywords", A1:D10)
 ```
 
 ```excel TORANGE
-=PROMPTMODEL.TORANGE("openai/gpt-4o-mini", "Extract keywords", A1:D10)
+=PROMPTMODEL.TORANGE("openai/gpt-5.4-mini", "Extract keywords", A1:D10)
 ```
 
 </ResponseExample>
diff --git a/docs/get-started/install.mdx b/docs/get-started/install.mdx
@@ -40,11 +40,11 @@ To install Cellm:
     If you want to use the more powerful hosted models right away, you can skip this step. The [Hosted Models](/models/hosted-models) section shows you how.
   </Info>
 
-  To get started with local models, we recommend you try out the Gemma 3 4B model with quantized aware training. Gemma 3 4B is a wonderful little model that will run fine on your CPU, ensuring no data ever leaves your computer. And it's free.
+  To get started with local models, we recommend you try out the Gemma 4 E4B model. Gemma 4 E4B is a wonderful little model that will run fine on your CPU, ensuring no data ever leaves your computer. And it's free.
 
   1. Download and install [Ollama](https://ollama.com/). Ollama will start after the install and automatically run whenever you start up your computer.
-  2. Open the Windows Terminal, type `ollama pull gemma3:4b-it-qat` and hit Enter.
-  3. Open Excel and type `=PROMPT("What model are you and who made you?")`. The model will respond that is is Gemma 3 and made by Google.
+  2. When you select an Ollama model in Cellm, it will prompt you to download it automatically. Alternatively, open the Windows Terminal, type `ollama pull gemma4:e4b` and hit Enter.
+  3. Open Excel and type `=PROMPT("What model are you and who made you?")`. The model will respond that it is Gemma 4 and made by Google.
 </Accordion>
 
 <Accordion icon="plug" title="Enable MCP (optional)">

diff --git a/docs/get-started/quickstart.mdx b/docs/get-started/quickstart.mdx
@@ -17,8 +17,8 @@ To get started, you can quickly install Cellm and a local model:
   <Step title="Install Ollama">
     Download and install [Ollama](https://ollama.com/) to run local AI models.
   </Step>
-  <Step title="Download Gemma 3 model">
-    Open the Windows Terminal and type `ollama pull gemma3:4b-it-qat` to download the Gemma 3 4B model.
+  <Step title="Download Gemma 4 model">
+    When you select an Ollama model in Cellm, it will prompt you to download it automatically. Alternatively, open the Windows Terminal and type `ollama pull gemma4:e4b` to download the Gemma 4 E4B model.
   </Step>
 </Steps>
 
@@ -143,7 +143,7 @@ Beyond basic text processing, you can use "Function Calling" to give models acce
 ```` 
 
 <Tip>
-  Gemma 3 4B does not support function calling. For function calling you must use another model, e.g. OpenAI's `gpt-5-mini`.
+  Gemma 4 E4B does not support function calling. For function calling you must use another model, e.g. OpenAI's `gpt-5.4-mini`.
 </Tip>
 
 ## Next steps

diff --git a/docs/models/choosing-model.mdx b/docs/models/choosing-model.mdx
@@ -91,12 +91,12 @@ Imagine you want to analyze customer feedback from column A. Instead of a single
 
 3. Extract Suggestions: In column D, use a Large model to analyze the feedback and suggest improvements. You could also add relevant background information on your product directly to the prompt or to a cell that you reference.
     ````mdx Analyze feedback
-    =PROMPTMODEL("openai/gpt-4o-mini", "Analyze user feedback and suggest improvements.", B2)
+    =PROMPTMODEL("openai/gpt-5.4-mini", "Analyze user feedback and suggest improvements.", B2)
     ````
 
 4. Extract Topics: In column E, extract relevant topics with a small model, which is efficient for simple extraction tasks.
     ````mdx Extract topics
-    =PROMPTMODEL.TOROW("openai/gpt-4o-mini", "Extract relevant software engineering topics, such as UX, Bug, Documentation, or Improvement.", B2)
+    =PROMPTMODEL.TOROW("openai/gpt-5.4-mini", "Extract relevant software engineering topics, such as UX, Bug, Documentation, or Improvement.", B2)
     ````
 
 This approach gives you reliable results and granular control of the output format.

diff --git a/docs/models/hosted-models.mdx b/docs/models/hosted-models.mdx
@@ -20,7 +20,7 @@ We split hosted models into three tiers based on their size and capabilities, ba
 | Speed             | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" />       | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" />   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" /> |
 | Intelligence      | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" />   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" />   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" />     |
 | World Knowledge   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" />   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" /> | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" />   |
-| Recommended model | Gemini 2.5 Flash Lite                                                                                                 | Gemini 2.5 Flash                                                                                                    | Claude Sonnet 4.5                                                                                                      |
+| Recommended model | Gemini 3.1 Flash Lite                                                                                                 | Gemini 3 Flash                                                                                                      | Claude Opus 4.6                                                                                                        |
 
 ## Provider setup
 
@@ -63,7 +63,7 @@ Mistral offers a generous free tier with access to powerful models.
 
 ### OpenAI
 
-OpenAI provides access to GPT models, including GPT-4o and GPT-4o-mini.
+OpenAI provides access to GPT models, including GPT-5.4 and GPT-5.4-mini.
 
 <Steps>
   <Step title="Create an account">
@@ -79,7 +79,7 @@ OpenAI provides access to GPT models, including GPT-4o and GPT-4o-mini.
     In Excel, open Cellm's ribbon menu, select the `openai` provider, click the provider icon, and paste your API key. Try a model like:
 
     ````mdx OpenAI example
-    =PROMPTMODEL("openai/gpt-4o-mini", "Classify sentiment as positive, neutral, or negative", A1)
+    =PROMPTMODEL("openai/gpt-5.4-mini", "Classify sentiment as positive, neutral, or negative", A1)
     ````
   </Step>
 </Steps>
@@ -106,7 +106,7 @@ Google Gemini offers powerful AI models with a generous free tier.
     In Excel, open Cellm's ribbon menu, select the `gemini` provider, click the provider icon, and paste your API key. Try a model like:
 
     ````mdx Gemini example
-    =PROMPTMODEL("gemini/gemini-2.5-flash", "Extract person names from text", A1)
+    =PROMPTMODEL("gemini/gemini-3-flash-preview", "Extract person names from text", A1)
     ````
   </Step>
 </Steps>
@@ -136,7 +136,7 @@ Anthropic provides Claude models, known for their strong reasoning capabilities.
     In Excel, open Cellm's ribbon menu, select the `anthropic` provider, click the provider icon, and paste your API key. Try a model like:
 
     ````mdx Claude example
-    =PROMPTMODEL("anthropic/claude-sonnet-4.5", "Analyze customer feedback", A1)
+    =PROMPTMODEL("anthropic/claude-sonnet-4-6", "Analyze customer feedback", A1)
     ````
   </Step>
 </Steps>

diff --git a/docs/models/local-models.mdx b/docs/models/local-models.mdx
@@ -16,7 +16,7 @@ We can split local models into three tiers based on their size and capabilities,
 | Speed             | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" />       | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" />   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" /> | 
 | Intelligence      | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" />   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" />   | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" />     | 
 | World Knowledge   | <Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" /> | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" /><Icon icon="star" iconType="regular" /> | <Icon icon="star" iconType="solid" /><Icon icon="star" iconType="solid" /><Icon icon="star" iconType="regular" />   |
-| Recommended model | Gemma 3 4B IT QAT                                                                                                     | Mistral Small 3.2                                                                                                   | qwen3-30b-a3b-instruct-2507                                                                                         |
+| Recommended model | Gemma 4 E4B                                                                                                            | Gemma 4 26B                                                                                                         | Gemma 4 31B                                                                                                          |
 
 <Tip>
 You need a GPU for any of the medium or large models to be useful in practice. If you don't have a GPU, you can use [Hosted Models](/models/hosted-models) if small ones are insufficient.
@@ -38,29 +38,29 @@ You need to run a program on your computer that serves models to Cellm. We call
 
 ### Ollama
 
-To get started with Ollama, we recommend you try out the Gemma 3 4B IT QAT model, which is Cellm's default local model.
+To get started with Ollama, we recommend you try out the Gemma 4 E4B model, which is Cellm's default local model.
 
 <Steps>
   <Step title="Install Ollama">
     Download and install [Ollama](https://ollama.com/). Ollama will start after the install and automatically run whenever you start up your computer.
   </Step>
   <Step title="Download the model">
-    Open Windows Terminal (open start menu, type `Windows Terminal`, and click `OK`), then run:
+    When you select an Ollama model in Cellm, it will prompt you to download it automatically. Alternatively, open Windows Terminal (open start menu, type `Windows Terminal`, and click `OK`), then run:
 
-    ````bash Download Gemma 3 4B QAT
-    ollama pull gemma3:4b-it-qat
+    ````bash Download Gemma 4 E4B
+    ollama pull gemma4:e4b
     ````
 
     Wait for the download to finish.
   </Step>
   <Step title="Test in Excel">
-    In Excel, select `ollama/gemma3:4b-it-qat` from the model dropdown menu, and type:
+    In Excel, select `ollama/gemma4:e4b` from the model dropdown menu, and type:
 
     ````mdx Test prompt
     =PROMPT("Which model are you and who made you?")
     ````
 
-    The model will tell you that it is called "Gemma" and made by Google DeepMind.
+    The model will tell you that it is called "Gemma 4" and made by Google DeepMind.
   </Step>
 </Steps>
 
@@ -133,7 +133,7 @@ If you prefer to run models via docker, both Ollama and vLLM are packaged up wit
     ````
   </Step>
   <Step title="Configure Cellm">
-    Start Excel and select the `openaicompatible` provider from the model drop-down on Cellm's ribbon menu. Enter the model name you want to use, e.g., `gemma3:4b-it-qat`.
+    Start Excel and select the `openaicompatible` provider from the model drop-down on Cellm's ribbon menu. Enter the model name you want to use, e.g., `gemma4:e4b`.
 
     Set the Base Address to `http://localhost:11434`.
   </Step>

diff --git a/docs/sdk-migration.md b/docs/sdk-migration.md
@@ -0,0 +1,33 @@
+# SDK Migration: Anthropic, Mistral, and Gemini
+
+## Summary
+
+Three provider SDKs need replacing due to incompatibilities and maintainability concerns. Both community SDKs (`Anthropic.SDK`, `Mistral.SDK`) are maintained by the same author (tghamm) and have become inactive. The Gemini provider uses an OpenAI-compatible endpoint that doesn't fully support tool use schemas.
+
+## ~~Anthropic.SDK (5.10.0)~~ ✅ DONE
+
+Migrated to official `Anthropic` SDK (v12.11.0). The community `Anthropic.SDK` was incompatible with MEAI 10.4.x (`MissingMethodException` on `HostedMcpServerTool.AuthorizationToken`). The official SDK has native IChatClient support and accepts custom HttpClient, so the resilient HttpClient pipeline is preserved. Also fixed a bug where the entitlement check referenced `EnableAzureProvider` instead of `EnableAnthropicProvider`. Removed the `RateLimitsExceeded` exception from `RateLimiterHelpers` (was Anthropic.SDK-specific; 429 status is already handled by `retryableStatusCodes`). All 4 integration tests pass.
+
+## ~~Mistral.SDK (2.3.1)~~ ✅ DONE
+
+Migrated both `AddMistralChatClient()` and `AddCellmChatClient()` to use `OpenAIClient` with custom endpoint, same pattern as DeepSeek and OpenRouter. Removed `Mistral.SDK` dependency entirely. All 4 Mistral integration tests pass (basic prompt, file reader, file search, Playwright MCP).
+
+**Known issue: Magistral thinking models.** The OpenAI .NET SDK cannot deserialize Magistral's `thinking` content part type (`ArgumentOutOfRangeException: Unknown ChatMessageContentPartKind value: thinking`). The failure occurs at the deserialization level before `MistralThinkingBehavior` can process the response. This is a limitation of using the OpenAI SDK with Mistral's extended thinking format. Magistral models (`magistral-small-2509`, `magistral-medium-2509`) are currently broken.
+
+## ~~Gemini (OpenAI-compatible endpoint)~~ ✅ DONE
+
+Migrated to official `Google.GenAI` SDK (v1.6.1). The OpenAI-compatible endpoint rejected `strict: true` / `additionalProperties: false` in tool schemas. The native SDK handles tool schemas correctly. All 4 integration tests pass (basic prompt, file reader, file search, Playwright MCP).
+
+**Tradeoff:** Google.GenAI does not support custom HttpClient injection, so HTTP-level retry/timeout from the resilience pipeline is not available for Gemini. Rate limiting (application layer) is unaffected. `GeminiTemperatureBehavior` (0-1 → 0-2 scaling) is still needed — the native SDK passes temperature as-is.
+
+## Additional considerations
+
+When switching SDKs, provider-specific behaviors and other code may need updating. Examples include but are not limited to:
+
+- `GeminiTemperatureBehavior` — temperature scaling may differ with native SDK
+- `AdditionalPropertiesBehavior` — provider-specific additional properties format may change
+- `ProviderRequestHandler.UseJsonSchemaResponseFormat()` — structured output support flags
+- Provider configuration classes (`SupportsJsonSchemaResponses`, `SupportsStructuredOutputWithTools`) — verify accuracy with new SDKs
+- Resilient HTTP client integration — new SDKs may handle HTTP clients differently
+
+A thorough review of all provider-specific code paths is needed during migration.
diff --git a/docs/usage/writing-prompts.mdx b/docs/usage/writing-prompts.mdx
@@ -104,7 +104,7 @@ For advanced workflows, you might want to use different AI models for different
 The first argument consists of a provider and a model name separated by a forward slash (`/`). For example, if you want to use OpenAI's cheapest model in a particular cell, you can write:
 
 ````mdx Specify model
-=PROMPTMODEL("openai/gpt-4o-mini", "Rate sentiment as positive, neutral, or negative", A1)
+=PROMPTMODEL("openai/gpt-5.4-mini", "Rate sentiment as positive, neutral, or negative", A1)
 ````
 
 This is useful when you want to use a strong model by default but offload simple tasks to cheaper models.