LLM Models · Reasoning
Browse the world's large language models. Compare parameters, benchmarks, VRAM and more.
Reasoning listings are models that spend extra tokens deliberating before the visible answer, or that are marketed for harder problems: multi-step bugs, proofs, long agent traces. The extra tokens are a cost and a latency, not a moral virtue. Use them when the task is actually hard. Do not use them to extract an email address from a signature.
You still need the ordinary cuts. A reasoning model with a short window cannot think its way through a repository that does not fit. A proprietary reasoning model still sends the prompt to the vendor. An open one still needs a GPU. How to read a model listing does not grow a special exception for this category. The thinking trace, when the API exposes it, is data too. Know whether it is stored.
Compare reasoning models on a problem you can grade. A puzzle the vendor published is a demo. A bug from your tracker, with the failing test attached, is an evaluation. Deciding after a comparison prefers the second. If both models solve it and one is half the price, keep the cheaper one unless the trace shows the cheap one got lucky.
Do not stack a reasoning model under a routing plugin that also "thinks" by calling three models, unless you have written down the bill. Inference plugins and reasoning models compound cost. Read model inference before you enable both.
A higher score on this shelf means the listing sorted higher. It does not mean the hidden chain of thought is inspectable or faithful. If you cannot see the trace and you cannot check the answer, you do not have a reasoning feature. You have a slower API.
Time the call. A reasoning model that is three times slower needs to be right more often, not once. If your graded set does not show that, turn the extra thinking off or pick the cheaper row. Latency is part of the product.
GPT-6.1 Sol is an upgrade to GPT-6 Sol from OpenAI, positioned below the flagship GPT-6 Astra in the GPT-6 series. It is suited for agentic coding, computer use, document-heavy professional...
GPT-6 Sol Pro is the same underlying model as [GPT-6 Sol](https://openrouter.ai/openai/gpt-6-sol), served with `reasoning.mode` set to `pro` for higher-quality responses on complex tasks. **Cost note:** pro mode spends far more...
:--- :--- Total Parameters 120B (12B active) Architecture LatentMoE - Mamba-2 + MoE + Attention hybrid with Multi-Token Prediction (MTP) Context Length Up to 1M tokens Minimum GPU Requirement 2× H100-80GB Supported Languages English, French, German, Italian, Japanese, Spanish, Ch…
Check out the NVFP4+GPTQ weights by FuriosaAI ! ➡️ link
DeepSeek R1 Distill Qwen2.5 Bakeneko 32B (rinna/deepseek-r1-distill-qwen2.5-bakeneko-32b)
DeepSeek R1 Distill Qwen2.5 Bakeneko 32B GPTQ int4 (rinna/deepseek-r1-distill-qwen2.5-bakeneko-32b-gptq-int4)
DeepSeek R1 Distill Qwen2.5 Bakeneko 32B GPTQ int8 (rinna/deepseek-r1-distill-qwen2.5-bakeneko-32b-gptq-int8)
DeepSeek R1 Distill Qwen2.5 Bakeneko 32B AWQ (rinna/deepseek-r1-distill-qwen2.5-bakeneko-32b-awq)
:--- :--- Total Parameters 120B (12B active) Architecture LatentMoE - Mamba-2 + MoE + Attention hybrid with Multi-Token Prediction (MTP) Context Length Up to 1M tokens Minimum GPU Requirement 8× H100-80GB Supported Languages English, French, German, Italian, Japanese, Spanish, Ch…
:--- :--- Total Parameters 120B (12B active) Architecture LatentMoE - Mamba-2 + MoE + Attention hybrid with Multi-Token Prediction (MTP) Context Length Up to 1M tokens Minimum GPU Requirement 1× B200 OR 1× DGX Spark Supported Languages English, French, German, Italian, Japanese,…
1. Summary 2. Evaluation 3. Training Details 4. How to Use 5. Intended Use 6. Limitations 7. Security and Responsible Use 8. Software 9. License 10. Acknowledgements 11. Citation
MiniCPM Tech Report MiniCPM Wiki(Chinese) GitHub Repo UltraData Online Demo