Pins matching tag:Reasoning

IBM releases Granite 3.2 with switchable chain-of-thought reasoning in the 8B and 2B Instruct models and its first official vision-language model, Granite Vision 3.2 2B, all under Apache 2.0.

Google releases an experimental Gemini 2.5 Pro, its first 2.5 model and a thinking model that scores 18.8% on Humanity's Last Exam and 63.8% on SWE-Bench Verified with a 1 million token context window.

IBM releases Granite 3.3 8B and 2B Instruct with fill-in-the-middle support and refined reasoning, plus Granite Speech 3.3 8B, its first official speech-to-text model, under Apache 2.0.

Google puts Gemini 2.5 Flash in preview, its first fully hybrid reasoning model, where developers can switch thinking off or cap it with a thinking budget of up to 24,576 tokens.

Alibaba's Qwen team open-sources the Qwen3 family under Apache 2.0, six dense models from 0.6B to 32B and two mixture-of-experts models (30B-A3B and 235B-A22B), with switchable thinking and non-thinking modes and 119 languages; the dense flagship Qwen3-32B scores 93.8 on Arena-Hard in thinking mode.

DeepSeek upgrades its R1 reasoning model to R1-0528 with open weights under the MIT license, lifting AIME 2025 from 70.0% to 87.5% and GPQA-Diamond from 71.5% to 81.0% while adding function calling and JSON output.

MiniMax releases M1, an open-weight 456B hybrid mixture-of-experts reasoning model with lightning attention, a 1M-token context and 80K-token reasoning output, trained with a new reinforcement-learning algorithm for a reported $534,700.

Alibaba's Qwen team releases Qwen3-235B-A22B-Thinking-2507, a thinking-only 235B mixture-of-experts model with a 262K context under Apache 2.0, scoring 92.3 on AIME25 and 74.1 on LiveCodeBench v6.

OpenAI releases GPT-5 mini and GPT-5 nano in the API with GPT-5: 400,000-token context, new minimal reasoning effort, and prices of $0.25 and $0.05 per 1 million input tokens.

Alibaba releases Qwen3-Next-80B-A3B-Instruct and Thinking, ultra-sparse 80B models that activate only 3B parameters per token, run a hybrid Gated DeltaNet attention design and carry an Apache 2.0 licence.

DeepSeek launches DeepSeek-V3.2, an open-weight model that thinks inside tool use and performs at GPT-5 level by its own tests, and the API-only DeepSeek-V3.2-Speciale, which it says rivals Gemini-3.0-Pro and reached gold-medal level at the 2025 IMO and IOI.

Google rolls out Gemini 3 Deep Think to Google AI Ultra subscribers in the Gemini app, an enhanced reasoning mode that scores 41.0% on Humanity's Last Exam and 45.1% on ARC-AGI-2.

Z.ai (Zhipu AI) releases GLM-4.7, an MIT-licensed coding-focused flagship that scores 73.8 on SWE-bench Verified and 87.4 on τ²-Bench, ranking first among open models on Code Arena at the end of 2025.

StepFun open-sources Step 3.5 Flash, a 196B-parameter sparse mixture-of-experts with 11B active parameters, a 256K context and up to 350 tokens per second, scoring 74.4% on SWE-bench Verified under Apache 2.0.

Alibaba releases Qwen3.5-397B-A17B, an open-weight native multimodal model that activates 17B of 397B parameters, plus the hosted Qwen3.5-Plus with a 1M-token context, on the eve of Lunar New Year.

Google releases Gemini 3.1 Pro in preview, an upgraded core for complex reasoning that reaches a verified 77.1% on ARC-AGI-2 and arrives in the Gemini app, NotebookLM, the API and Vertex AI.

Inception releases Mercury 2, which it calls the first reasoning diffusion LLM: 1,009 tokens per second on NVIDIA Blackwell GPUs at $0.25 per 1 million input tokens and $0.75 per 1 million output tokens, with a 128K context.

Arcee AI releases Trinity-Large-Thinking, a 398B-parameter open reasoning model (about 13B active) built for long-horizon agents, which it says ranks second on PinchBench behind Claude Opus 4.6 at $0.90 per 1 million output tokens.

Alibaba releases Qwen3.6-Plus, a hosted multimodal model aimed at agentic coding with a 1M-token context window, priced at CNY 2 per 1M input tokens, scoring 78.8 on SWE-bench Verified.

SpaceXAI (formerly xAI) puts Grok 4.3 on its API, a 1M-token reasoning model at $1.25 per 1 million input tokens and $2.50 per 1 million output tokens, about 40% and 60% cheaper than Grok 4.20; Artificial Analysis scored it 53 at launch.

Alibaba unveils Qwen3.7-Max, a proprietary text-only flagship agent model with a 1M-token context window that it says ran autonomously for 35 hours and scores 80.4 on SWE-bench Verified.

Tencent officially releases Hy3, the production version of its open-weight Hunyuan model: a 295B-parameter mixture-of-experts with 21B active parameters and a 256K-token context, priced at 1 yuan per 1 million input tokens.

Z.ai (Zhipu AI) releases GLM-5.3, a post-training-only upgrade of GLM-5.2 that scores 28.3 on Terminal Bench 3.0 and 84.5% on the CyberGym vulnerability benchmark, with a 1M-token context and open weights two weeks later.

IBM releases Granite 4.2, three open Apache 2.0 reasoning models (3B, 8B and 30B) with a switchable thinking mode; the 8B scores 47.67 on SWE-bench Verified and 86.67 on AIME25 in IBM's own runs.