Search “Reasoning”

Thinking Machines releases Inkling 2 Small, which scores 31 on the Artificial Analysis Intelligence Index v5.0.1, 18 points higher than Inkling Small and on par with Gemini 3.8 Flash (high). The model is currently proprietary, offered through the Tinker platform and Tinker Playground, with plans to later open its weights under the Apache 2.0 license. The model has 594B total parameters and 17B activated parameters, supports text and image input, and is priced at $0.30/$1.00 per 1M input/output t

Matched fieldsBody

Cognition has released the SWE-2 Flash coding model, trained on Thinking Machines' Inkling-Small (276B total parameters, 12B activated parameters), marking the first time a mid-training stage before reinforcement learning was added to the training pipeline. The company claims Terminal-Bench 4.0 improved from 0.9% to 20.0%, FrontierCode 2.0 from 3.2% to 24.0%, and ExploitBench from 11.3% to 34.8%. The model is now available on Devin Desktop and the CLI, priced at one-fifth of its larger sibling S

Matched fieldsBody

Matched fieldsBody

Under the Cosmos world model platform, NVIDIA has added four Cosmos 3 open-weight checkpoints, all post-trained versions of Cosmos 3 Nano, containing 16B trainable parameters and released under the OpenMDW1.1 license, which permits commercial and non-commercial use. The four models respectively target multi-view driving data generation, synchronized streaming simulation of cameras and LiDAR, simulator-controlled generation of photorealistic video, and video prediction from dual-arm robot actions

Matched fieldsBody

OpenAI releases the GPT-6 Cyber system card, a cybersecurity-specialized variant of GPT-6 Astra intended for authorized security work such as vulnerability discovery, exploit development, and web penetration testing. In defensive-use evaluations, the model completed 100% of general security assistance and vulnerability remediation tasks, 96% of exploit development tasks, and 97% of red-team tasks. Under the Preparedness Framework, its capability ratings are the same as GPT-6 Astra: cybersecurity

Matched fieldsBody

Tencent Hunyuan Releases Hy4 Official Version, Significantly Expanding Pre-training and Post-training Scale Compared with Hy4 preview, Introducing Vision Capabilities and Supporting Computer Operation for the First Time. The company says Hy4 has for the first time ranked among the strongest open-source models, improving by 19.1% and 9.8% on the GameDev and GameLogic benchmarks respectively, while the hallucination rate dropped from 15.76% to 5.45%. Hy4 is now available on Tencent Cloud and OpenR

Matched fieldsBody

Sarvam AI has released Sarvam Prism, a System One decision model based on Sarvam 105B, which achieves an average score of 71.8 with a median response time of 44 ms on B200 GPUs, the highest score within 100 ms among the 25 decision models it tested. At confidence levels above 90%, the model's actual error rate is lower than expected: 2.2% versus 3.3% for English and 4.4% versus 5.4% for Indic languages, and it keeps its answers unchanged in a 2,000-question option-shuffling test. The weights are

Matched fieldsBody

French company Mistral AI has released the Large 4 model (ML4), featuring advanced coding and cyber defense capabilities; enterprises can download the open-weight version and deploy it to run on their own infrastructure. The model has a total of 1.05 trillion parameters, with approximately 52 billion parameters actually activated per task execution, and was trained in a European data center using 3,800 NVIDIA Grace Blackwell GPUs. Mistral claims it is comparable to other leading open models on c

Matched fieldsBody

Edge0 has released the open-weight voice cloning TTS model Audio8-TTS-Realistic, with 601M parameters, supporting 23 languages, with weights and code released under Apache 2.0. The model requires only 1 reference audio clip and its corresponding text for zero-shot voice cloning, and can also directly generate speech when no reference voice is available. Self-reported results show an average error rate of 4.86% on the CV3 nine-language benchmark, lower than Qwen3-TTS-1.7B's 7.73%, but it lags beh

Matched fieldsBody

Google has released its new-generation flagship model Gemini 4 Carbon, which tops the Artificial Analysis Intelligence Index v5.0 with 59 points, surpassing Claude Opus 5.5 and GPT-6 Astra. On Terminal-Bench 4.0, the model improved from Argon's 57.1% to 65.2%, and on the GraphWalks long-context test it rose from 84.2% to 96.2%, but developers report that its generation speed is about 2 to 4 times slower than Argon. This week Google first opened it to a small number of cybersecurity defense organ

Matched fieldsBody

Ant Group's GuangJian team has open-sourced DefakerOne2 on GitHub, a 2B-parameter unified visual forensics model built on the Qwen3.5-2B multimodal backbone, covering AI-generated content, DeepFake faces, document tampering, natural image tampering, face and ID anti-counterfeiting, and AI-generated video detection. The model was trained on approximately 50 million samples, employs a three-stage Tide-Style Tuning process from LoRA to full parameters and back to LoRA, and uses entropy-based test-t

Matched fieldsBody

Baseten's Base Labs has released joey-1, a 2B decision model that directly outputs a single-token label and calibrated confidence, with a P50 end-to-end latency of 20ms, about 2.5 times faster than elephant-1 (49ms), at a price of $0.01 per million tokens, one-tenth that of mammoth-1. joey-1 is post-trained from Qwen3.5-2B and shares the same data recipe and API with elephant-1 and mammoth-1, so switching requires changing only one line. The model is now available on Baseten Model APIs, with zer

Matched fieldsBody

Matched fieldsBody

NVIDIA has released the open-weight model Nemotron 3.5 Transcribe on Hugging Face and ModelScope, integrating transcription, speaker diarization, translation, and audio question answering into a single checkpoint. It supports transcription in 40 languages, translation between 31 languages and English, and speaker labeling in 13 languages, can process recordings up to 1 hour long, and can label up to 8 speakers. The model ranks first among open-weight models on the English leaderboard of the Hugg

Matched fieldsBody

Mistral AI updated and opened the weights of its flagship model Mistral Large 4, with 1.05T total parameters and 52B activated parameters. It scores 26 on the Artificial Analysis Intelligence Index v5, 1 point higher than the Preview version, and leads the US open-source model Muse Spark Open (25) by 1 point. The model outputs 82K tokens per task, 39% fewer than the Preview's 135K, and its hallucination rate on AA-Omniscience v1.1 dropped from 41% to 28%, though accuracy remains 26%. Pricing is

Matched fieldsBody

Beijing-based large model startup Chaoyan Intelligence has released and open-sourced its first large model, Apex-35B-V1.0, with 35 billion total parameters, approximately 3 billion parameters activated per token, and support for a 262,000-token context window, optimized through mid-training and post-training based on Qwen3.6-35B-A3B. The model scored 60.9 and 40.0 on the paper reproduction test PaperBench and the scientific research challenge test FrontierScience, respectively, higher than GLM-5

Matched fieldsBody

On October 28, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that can transcribe audio in 40 languages, translate speech, and identify speakers, and developers can control the transcription format with natural language instructions without retraining. The model consists of a Transformer speech encoder, a speaker diarization model, and the Nemotron 3.5 Lightning language model, with approximately 4.3 billion activated parameters, and the released checkpoint is labeled 30B-A

Matched fieldsBody

Google DeepMind launched Gemini 4 Carbon two weeks after the release of Argon. The model scored 59 on the Artificial Analysis Intelligence Index v5.0, higher than Argon's 50, ranking first. Carbon is designed for long-horizon engineering tasks, rose to 1769 points and fourth place on WebDev Arena, and reduced the attack success rate from 0.7% to 0.0% on the Gray Swan indirect prompt injection benchmark. The model is first being made available to select cybersecurity experts, priced at $2 per mil

Matched fieldsBody

Mistral has released the weights of Mistral Large 4 on Hugging Face. The model, nicknamed Le Chonk, has approximately 1 trillion parameters with 52 billion activated per step, and supports text and image inputs. On Terminal-Bench 4.0, its score rose from 28% in the preview version to 41%, slightly above GLM-5.3's 40%, but on the new version of the Artificial Analysis Intelligence Index it scored only 26 points, below the 32 points of DeepSeek V4.1 Flash and GLM-5.3. API pricing is $0.70 per mill

Matched fieldsBody

Google DeepMind has released Gemini 4 Carbon, which ranks first on the Artificial Analysis Intelligence Index v5 with a score of 59 and leads in 4 of 9 evaluations including HLE-Diamond and TB-Science. The model focuses on cybersecurity and long-horizon tasks, claiming 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench, and is initially available only to trusted cybersecurity personnel through the Fairwind Program. In terms of specifications, it supports a 1M token context and 1

Matched fieldsBody