> ## Documentation Index
> Fetch the complete documentation index at: https://documentation.deepmask.io/llms.txt
> Use this file to discover all available pages before exploring further.
# Choosing Right Model
> Match your task to the right AI model in DeepMask. Compare models by use case — coding, research, writing, speed, reasoning, and EU-only data residency.
DeepMask is built around the principle that no single model is best for every task. You can switch models at any point — including mid-conversation — without losing your context. Use the tabs below to find the right model for what you're working on right now.
These models perform well on software engineering tasks including code generation, debugging, architecture design, and long-horizon autonomous development.
| Model | Why it fits |
| ----------------------- | ------------------------------------------------------------------------------------------------ |
| **GPT-5.2 / 5.3 / 5.4** | Most capable OpenAI models; strong reasoning and tool use across all coding tasks |
| **Sonnet 4.5 / 4.6** | Gold standard for autonomous coding; handles 30+ hour engineering sessions with 1M token context |
| **Gemini 2.5 Pro** | Large context window suits large codebase analysis and multi-file refactors |
| **Kimi K2 (DeepMask)** | Agent Swarm Mode enables 100 parallel sub-agents for complex, multi-step builds |
| **DeepSeek V3** | 671B MoE model with frontier-level coding and math; strong on STEM and security analysis |
| **MiniMax M2 / M2.1** | Built specifically for elite multi-language coding and advanced agent workflows |
These models excel at synthesizing large volumes of information, reasoning over documents, and producing structured analytical outputs.
| Model | Why it fits |
| ----------------------- | ----------------------------------------------------------------------------------------------------------- |
| **Kimi K2 (DeepMask)** | Searches hundreds of sources simultaneously in Agent Swarm Mode; 2M token context for massive document sets |
| **Opus 4.5 / 4.6** | Anthropic's highest-capability tier; strong on complex multi-document reasoning and subtle logical analysis |
| **GPT-5.2 / 5.3 / 5.4** | Full document and image analysis with strong reasoning across all content types |
| **DeepSeek V3** | Outperforms most frontier models on AIME and MATH-500; strong for STEM research and symbolic math |
| **Gemini 2.5 Pro** | 1M+ token context; suited for summarizing large research corpora |
| **Gemini 2.5 Flash** | Handles real-time summarization of hundreds of PDFs or hour-long recordings in one pass |
These models produce high-quality long-form text, adapt to different tones, and handle creative and professional writing tasks.
| Model | Why it fits |
| -------------------- | ------------------------------------------------------------------------------------------------------------------- |
| **GPT-5.2** | Highly capable for creative and persuasive writing; shown in-product generating marketing copy and strategy content |
| **Opus 4.5** | Nuanced, high-fidelity writing with strong narrative coherence; suited for strategy, legal, and financial documents |
| **Mistral Large 3** | Elite multilingual writing across 40+ languages; good for international marketing and professional content |
| **Sonnet 4.5** | Balanced between quality and speed; well-suited for content workflows requiring document context |
| **Mistral Medium 3** | Frontier-level writing output at significantly lower cost; good for high-volume content generation |
When you need quick responses, high throughput, or a cost-efficient model for simple queries and automation pipelines, these models deliver.
| Model | Why it fits |
| -------------------- | -------------------------------------------------------------------------------------------------------- |
| **Haiku 4.5** | Anthropic's fastest model at 180+ tokens/sec with 0.20s latency; designed for enterprise-scale workloads |
| **Gemini 2.5 Flash** | 185 tokens/sec with a 1M token context window; most cost-effective for large-volume document processing |
| **GLM-4.7 Flash** | Lightweight MoE model with strong reasoning and coding accuracy at high speed |
| **Mistral Medium 3** | 8× lower cost than frontier-tier models with strong general performance |
These models apply extended or structured thinking to work through complex, multi-step problems including math, logic, planning, and ambiguous tasks.
| Model | Why it fits |
| ----------------------- | ------------------------------------------------------------------------------------------------------------ |
| **Kimi K2 (DeepMask)** | High reasoning effort; 87.6% GPQA Diamond; decomposes tasks into 100 parallel sub-tasks |
| **Sonnet 4.5 / 4.6** | Adaptive reasoning (standard/high); 83.4% GPQA Diamond; significant gains on graduate-level math and science |
| **GPT-5.2 / 5.3 / 5.4** | Strong reasoning built into the latest GPT-5 generation |
| **GPT-o3 Mini** | Dedicated reasoning model for document analysis and tool use |
| **DeepSeek V3** | Adaptive non-thinking/thinking mode; 80.7% GPQA Diamond; excellent for mathematical proofs |
| **GLM-4.7** | Interleaved thinking with elite agent workflows; strong on complex real-world tasks |
If your organization requires that all data processing occurs within the European Union, every model in this tab runs exclusively on EU infrastructure.
| Model | Host | Notes |
| ---------------------------- | -------------------------------- | ------------------------------------------------- |
| **Qwen3 (StackIT)** | StackIT (Schwarz Group, Germany) | Reasoning, tool use, document and image analysis |
| **GPT-OSS 120B (StackIT)** | StackIT (Schwarz Group, Germany) | Document analysis and research; no image support |
| **Gemma 3 27B (StackIT)** | StackIT (Schwarz Group, Germany) | Lightweight; chat and document/image analysis |
| **DeepSeek V3.1 (Infercom)** | Infercom (EU-hosted endpoints) | Document analysis, tool use, complex writing |
| **GPT-OSS 120B (Infercom)** | Infercom (EU-hosted endpoints) | Document analysis and research; no image support |
| **MiniMax M2.5 (Infercom)** | Infercom (EU-hosted endpoints) | Coding, document analysis, tool use; 164K context |
| **Kimi K2 (DeepMask)** | DeepMask EU infrastructure | Full capability including Agent Swarm Mode |
| **Qwen (DeepMask)** | DeepMask EU infrastructure | Reasoning, tool use, multilingual chat |
StackIT is operated by the Schwarz Group (parent company of Lidl and Kaufland) and is certified as a German sovereign cloud. Infercom provides EU-hosted LLM endpoints with strict data residency controls. DeepMask's own infrastructure is also EU-based.
You can switch models at any point in a conversation. If a response isn't working well, select a different model from the model picker and continue your conversation — DeepMask carries your context forward automatically.
## Model capability quick reference
Extended thinking (sometimes shown as "reasoning mode" or "thinking mode" in the UI) causes a model to work through a problem step-by-step before producing its final answer. The model generates an internal chain of reasoning that it uses to improve accuracy on complex tasks.
Models in DeepMask that support extended or adaptive thinking include **Sonnet 4.5 / 4.6**, **Haiku 4.5**, **Kimi K2 (DeepMask)**, **GLM-4.7**, and **DeepSeek V3** (in its thinking mode). GPT-o3 Mini is also specifically optimized for reasoning tasks.
Extended thinking increases response time but significantly improves results for graduate-level math, multi-step logic, planning, and any task where intermediate reasoning matters.
The following models in DeepMask can accept images as input and reason about their contents:
* **OpenAI:** GPT-4o, GPT-4.1, GPT-5.2, GPT-5.3, GPT-5.4
* **Anthropic:** Opus 4.5 / 4.6, Sonnet 4.5 / 4.6, Haiku 4.5
* **Google:** Gemini 2.5 Pro, Gemini 2.5 Flash, Gemma 3 27B (StackIT)
* **MoonshotAI:** Kimi K2 (DeepMask), Kimi K2.5 (via MoonViT multimodal)
* **Alibaba:** Qwen (DeepMask), Qwen3 (StackIT)
* **Mistral:** Mistral Large 3
Models that do **not** support image input include GPT-OSS 120B (both StackIT and Infercom variants), DeepSeek V3 / V3.1, GPT-o3 Mini, and MiniMax M2.5 (Infercom).
Tool use (also called function calling) lets a model invoke external tools, APIs, or data connectors during a conversation. DeepMask exposes tool use through its MCP (Model Context Protocol) connector framework, which supports Google Drive, Gmail, SharePoint, Salesforce, and more.
Models with strong tool use support include:
* **Anthropic:** Haiku 4.5 (95% success rate on complex JSON schemas), Sonnet 4.5 / 4.6, Opus 4.5 / 4.6
* **OpenAI:** GPT-4o, GPT-4.1, GPT-5.x, GPT-o3 Mini
* **MoonshotAI:** Kimi K2 (DeepMask) — maintains coherence across 300+ sequential tool calls
* **Google:** Gemini 2.5 Pro, Gemini 2.5 Flash
* **Alibaba:** Qwen (DeepMask), Qwen3 (StackIT)
* **DeepSeek:** DeepSeek V3, DeepSeek V3.1 (Infercom)
* **MiniMax:** M2, M2.1, M2.5 (Infercom)
* **Z.ai:** GLM-4.7
Gemma 3 27B (StackIT) does not support tool use in DeepMask.
The context window is the maximum amount of text (measured in tokens, where 1 token ≈ 0.75 words) that a model can read and reason over in a single conversation. Larger context windows let you work with longer documents, more conversation history, and bigger codebases without losing earlier information.
Context window sizes for key models in DeepMask:
| Model | Context window |
| ----------------------- | ------------------------ |
| Kimi K2 (DeepMask) | 2,000,000 tokens |
| Sonnet 4.5 / 4.6 | 1,000,000 tokens |
| Gemini 2.5 Flash | 1,040,000 tokens |
| Gemini 2.5 Pro | \~1,000,000 tokens |
| Haiku 4.5 | 200,000 tokens |
| DeepSeek V3 / V3.1 | 128,000 – 164,000 tokens |
| MiniMax M2.5 (Infercom) | 164,000 tokens |
| GPT-4o | 128,000 tokens |
| GPT-4.1 | 128,000 tokens |
For very long documents or multi-session projects, prefer Kimi K2, the Sonnet series, or the Gemini 2.5 models. Use DeepMask Projects to persist files and instructions across sessions regardless of the model you choose.