Extended Reasoning With Thinking Tokens

via “thinking-models-and-extended-reasoning-support”

Get up and running with Kimi-K2.5, GLM-5, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

Unique: Thinking token handling is integrated into the inference pipeline, not a post-processing step. KV cache management accounts for thinking token overhead, preventing OOM errors when reasoning tokens exceed output tokens by orders of magnitude.

vs others: More transparent than OpenAI's o1 API because thinking tokens are accessible for debugging; more flexible than vLLM because it supports arbitrary thinking token formats without requiring model-specific parsing

Google's multimodal API — Gemini 2.5 Pro/Flash, 1M context, video understanding, grounding.

Unique: Allocates hidden 'thinking tokens' for internal reasoning before generating output, allowing the model to spend additional computation on difficult problems without exposing reasoning steps to the user

vs others: Similar to OpenAI's o1 extended reasoning, but integrated into the standard Gemini API rather than a separate model, allowing extended reasoning on the same multimodal inputs (images, audio, video) that standard Gemini supports

via “extended thinking with user-controlled reasoning effort”

Anthropic's balanced model for production workloads.

Unique: Implements hybrid reasoning with both user-controlled extended thinking and automatic adaptive thinking, allowing fine-grained effort control via API parameters rather than binary on/off toggle. This dual-mode approach enables cost optimization by letting developers choose reasoning depth per-request while maintaining automatic reasoning for complex queries.

vs others: Offers more granular reasoning control than GPT-4o's reasoning mode (which lacks effort parameters) and lower cost than o1 models while maintaining competitive reasoning performance on complex tasks.

via “reasoning and step-by-step problem decomposition”

text-generation model by undefined. 95,66,721 downloads.

Unique: Emergent chain-of-thought capability from instruction tuning on reasoning datasets; no explicit reasoning module or symbolic engine — reasoning emerges from learned token prediction patterns that favor intermediate explanation tokens, making it lightweight but probabilistic

vs others: Provides transparent reasoning comparable to GPT-4 on simple problems but with full local control; outperforms Mistral-7B on reasoning tasks due to instruction tuning, but lacks the formal verification and symbolic reasoning of specialized tools like Wolfram Alpha

via “extended thinking and reasoning mode for complex problem-solving”

Anthropic's developer console for Claude API.

Unique: Provides access to Claude's internal reasoning process via thinking blocks, allowing developers to inspect and debug Claude's reasoning rather than only seeing final outputs

vs others: More transparent than black-box reasoning in other LLMs, and allows developers to tune reasoning effort via budget parameters

via “extended-thinking-transparent-reasoning”

Anthropic's most intelligent model, best-in-class for coding and agentic tasks.

Unique: Separates thinking tokens from output tokens in the API response, allowing clients to inspect, log, or discard reasoning steps independently. This architectural choice enables cost-aware reasoning allocation — users can trade latency and cost for reasoning depth on a per-request basis, unlike competitors who bundle reasoning into standard inference.

vs others: More transparent and controllable than OpenAI o1's opaque reasoning, and more cost-granular than competitors by separating thinking token accounting from output tokens, enabling selective reasoning on high-complexity queries only.

via “chain-of-thought-multi-stage-reasoning”

Google's vision-language-action model for robotics.

Unique: Integrates chain-of-thought reasoning directly into the action generation pipeline by representing both reasoning steps and actions as text tokens, allowing the same transformer to generate interpretable intermediate steps and grounded robot actions

vs others: Provides interpretability and reasoning transparency that black-box policy networks lack, while avoiding separate symbolic reasoning systems by leveraging the language model's native ability to generate and process reasoning text

via “reasoning and extended thinking with streaming token consumption”

Unique: Void's reasoning support is model-aware; it detects reasoning capabilities in the LLM Message Service and configures provider-specific parameters (e.g., budget_tokens for Claude, reasoning effort for OpenAI). Thinking tokens are streamed separately from response tokens, allowing the UI to display the AI's reasoning process in real-time.

vs others: Unlike standard Copilot or Cursor (which use fast models without reasoning), Void's reasoning support enables deep thinking for complex problems, making it suitable for architectural decisions and code reviews where reasoning quality matters more than speed.

via “reasoning model output parsing with thinking extraction”

OpenAI and Anthropic compatible server for Apple Silicon. Run LLMs and vision-language models (Llama, Qwen-VL, LLaVA) with continuous batching, MCP tool calling, and multimodal support. Native MLX backend, 400+ tok/s. Works with Claude Code.

Unique: Parses and separates thinking tokens from final output during streaming, enabling real-time access to model reasoning without waiting for generation completion; supports multiple reasoning formats with configurable parsing strategies

vs others: More transparent than black-box reasoning (exposes thinking process); enables streaming reasoning display unlike batch-only parsing; supports multiple model formats

via “extended-reasoning-with-thinking-tokens”

Gemini 2.5 Pro is Google’s state-of-the-art AI model designed for advanced reasoning, coding, mathematics, and scientific tasks. It employs “thinking” capabilities, enabling it to reason through responses with enhanced accuracy...

Unique: Uses hidden thinking tokens that consume inference budget but remain invisible to users, enabling internal verification and multi-path exploration without exposing intermediate steps — distinct from chain-of-thought which exposes all reasoning to the user

vs others: Provides higher accuracy on complex reasoning tasks than standard LLMs while maintaining clean output formatting, though at higher latency and token cost than models without extended thinking capabilities

via “extended reasoning with native thinking mode”

Gemini 2.5 Flash is Google's state-of-the-art workhorse model, specifically designed for advanced reasoning, coding, mathematics, and scientific tasks. It includes built-in "thinking" capabilities, enabling it to provide responses with greater...

Unique: Integrates reasoning as a first-class inference primitive rather than a prompt engineering technique, using an internal thinking phase that explores solution spaces before output generation, with separate token accounting for transparency

vs others: Provides more reliable reasoning than prompt-based CoT approaches (like o1-preview) while maintaining faster inference than full-chain reasoning models, with explicit visibility into thinking token usage

via “extended-reasoning-with-internal-thinking”

Gemini 2.5 Pro is Google’s state-of-the-art AI model designed for advanced reasoning, coding, mathematics, and scientific tasks. It employs “thinking” capabilities, enabling it to reason through responses with enhanced accuracy...

Unique: Implements internalized thinking as part of the inference architecture rather than exposing chain-of-thought tokens, allowing the model to reason without token overhead while maintaining response quality. Uses adaptive computation allocation to balance reasoning depth with response latency based on problem complexity.

vs others: Provides reasoning benefits of extended chain-of-thought without the token cost and latency of explicit reasoning tokens, differentiating it from models like o1 that expose reasoning in the output stream.

via “extended-chain-of-thought reasoning with explicit thinking tokens”

Qwen3-Max-Thinking is the flagship reasoning model in the Qwen3 series, designed for high-stakes cognitive tasks that require deep, multi-step reasoning. By significantly scaling model capacity and reinforcement learning compute, it...

Unique: Uses dedicated thinking token architecture with RL-optimized allocation strategy, allowing the model to dynamically determine reasoning depth per query rather than applying fixed reasoning budgets like some competitors. Separates internal deliberation from output generation at the token level, enabling transparent reasoning traces.

vs others: Provides deeper, more transparent reasoning than standard LLMs while maintaining faster inference than some reasoning-specialized models by using learned heuristics to allocate thinking compute only when needed.

via “extended-reasoning-with-thinking-tokens”

Claude 3.7 Sonnet is an advanced large language model with improved reasoning, coding, and problem-solving capabilities. It introduces a hybrid reasoning approach, allowing users to choose between rapid responses and...

Unique: Implements server-side thinking tokens that are not billed to the user (or billed separately) and remain invisible to the client, enabling the model to perform exploratory reasoning without exposing intermediate steps. This differs from other CoT approaches (like OpenAI's o1) which may return reasoning traces or charge for all reasoning compute at the same rate as output.

vs others: Offers reasoning-enhanced responses at lower cost than o1-class models while maintaining faster latency than full reasoning models, positioned as a middle ground between fast-but-shallow Sonnet and slow-but-deep reasoning specialists.

via “extended reasoning with chain-of-thought for complex visual tasks”

Qwen3-VL-30B-A3B-Thinking is a multimodal model that unifies strong text generation with visual understanding for images and videos. Its Thinking variant enhances reasoning in STEM, math, and complex tasks. It excels...

Unique: Integrates extended reasoning directly into the model's forward pass for visual tasks, rather than using post-hoc prompting techniques like 'think step-by-step', enabling the model to allocate compute dynamically to reasoning-heavy visual problems

vs others: More reliable than prompt-based chain-of-thought for visual reasoning because reasoning is baked into model weights, not dependent on prompt engineering; produces more consistent intermediate steps for STEM tasks

via “extended reasoning with implicit chain-of-thought”

Grok 4 is xAI's latest reasoning model with a 256k context window. It supports parallel tool calling, structured outputs, and both image and text inputs. Note that reasoning is not...

Unique: Implicit reasoning allocation based on problem complexity, with reasoning traces integrated into output without explicit token budget management, contrasting with OpenAI's explicit reasoning token approach

vs others: More transparent reasoning than GPT-4o (which hides reasoning) but less controllable than o1 (which offers explicit reasoning token budgets); better for exploratory reasoning where depth is problem-dependent

via “hybrid-reasoning-with-explicit-thinking-mode”

DeepSeek-V3.1 is a large hybrid reasoning model (671B parameters, 37B active) that supports both thinking and non-thinking modes via prompt templates. It extends the DeepSeek-V3 base with a two-phase long-context...

Unique: Implements user-controlled explicit thinking via prompt templates rather than always-on reasoning, allowing per-request cost-performance optimization. The 37B active parameter subset processes thinking tokens in a separate phase before final generation, unlike models that interleave reasoning throughout decoding.

vs others: Offers finer-grained reasoning control than OpenAI o1 (which always reasons) and better cost efficiency than Claude 3.5 Sonnet's extended thinking by letting developers opt-in only when needed.

via “long-context reasoning with extended thinking”

Claude Opus 4.5 is Anthropic’s frontier reasoning model optimized for complex software engineering, agentic workflows, and long-horizon computer use. It offers strong multimodal capabilities, competitive performance across real-world coding and...

Unique: Implements internal chain-of-thought reasoning within a 200K token window using transformer attention mechanisms, allowing reasoning to occur before output generation without requiring explicit prompt engineering for step-by-step thinking

vs others: Outperforms GPT-4o and Claude 3.5 Sonnet on complex reasoning tasks by maintaining coherence across longer reasoning chains while keeping the 200K context window practical for real-world applications

via “reasoning-augmented text generation with explicit thinking mode”

Qwen3-8B is a dense 8.2B parameter causal language model from the Qwen3 series, designed for both reasoning-heavy tasks and efficient dialogue. It supports seamless switching between "thinking" mode for math,...

Unique: Implements explicit thinking mode as a native architectural feature rather than prompt-engineering workaround, using token-level gating to separate reasoning computation from response generation within a single 8B parameter model

vs others: Achieves reasoning performance comparable to 70B+ models while maintaining 8B parameter efficiency through dedicated thinking tokens, unlike Llama or Mistral which require larger model sizes or external chain-of-thought prompting

via “reasoning and chain-of-thought decomposition”

Mistral Large 2 2411 is an update of [Mistral Large 2](/mistralai/mistral-large) released together with [Pixtral Large 2411](/mistralai/pixtral-large-2411) It provides a significant upgrade on the previous [Mistral Large 24.07](/mistralai/mistral-large-2407), with notable...

Unique: Mistral Large 2411 implements implicit chain-of-thought through training on reasoning-heavy datasets, enabling natural step-by-step decomposition without explicit prompting while maintaining efficiency through optimized token generation

vs others: Provides reasoning quality comparable to GPT-4 while maintaining lower latency and cost through more efficient token usage