Search “Agents”

Google DeepMind releases Gemini 4 Carbon, focused on cybersecurity defense and coding

Google DeepMind has launched Gemini 4 Carbon, calling it a frontier model for long-horizon agentic coding and cybersecurity defense, ranking first with 59 points on the Artificial Analysis Intelligence Index v5.0. The model achieved 96.2% on the GraphWalks long-context test, and 88.2%, 54.8%, and 71.0% on SEC-bench Pro, ExploitGym, and CWE-bench, respectively, but on OSWorld-2.1 it trailed Claude Opus 5.5's 87.2% with 81.5%. Carbon is currently available only to vetted cybersecurity defenders, with introductory pricing of $2 per million input tokens and $10 per million output tokens, later $4 and $20; the timeline for opening it to developers and the public has not been announced.

OpenBMB Open-Sources MiniCPM-OCR Suite, Featuring a 1B Parsing Model and a 2B Verification Model

OpenBMB has released the open-source MiniCPM-OCR suite, comprising the 1B MiniCPM-OCR 1.0 and the 2B decision model MiniCPM-OCR Jev. MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2; Jev can perform cross-model verification of text, formulas, and tables, scoring 93.6 on JevOCRBench, surpassing JEV-35B-A3B's 90.1 and approaching Qwen3.5-397B-A17B's 93.8. Powered by DSpark, its speed is 1.49 times that of PaddleOCR-VL-1.6 and 3.62 times that of MinerU2.5 Pro; with SDK concurrency optimization, it can reach 128 pages per minute, and it supports deployment on vLLM, SGLang, llama.cpp, and Ollama.

Google DeepMind releases the world model Genie 4, natively generating sound and dialogue for the first time

Google DeepMind has released Genie 4, a new-generation general-purpose world model. For the first time in the Genie series, it natively generates ambient sound and improvised character dialogue, doubles the duration of a single coherent world to about 2 minutes, and renders real-machine footage at about 24 frames per second. The model uses Gemini 3.8 Flash to perform multi-step reasoning and control in the background and Gemini 4 Argon to break down action instructions, forming a hybrid world model architecture. It has opened beta applications to Google AI Ultra users and is also available in the Gemini API and Google AI Studio.

Liquid AI Releases the LFM3 Series of Models, Including 1.7B Text, Vision, Audio, and Omni Versions

Liquid AI has released the LFM3 series of open-weight models, all branched from LFM3-1.7B-Base, covering text, vision, audio, Omni, and the d1 decision model. LFM3-1.7B-Base adopts a three-hybrid-layer architecture, is pretrained on 22T tokens, and in the mid-training stage over the final 3.5T tokens incorporates more agent trajectories and tool-calling data, with the context window extended to 256K tokens. The company says LFM3-1.7B is about 25% faster than LFM2-1.2B at an 8K context, and scores 38.42 versus 3.02 on WebArena-Verified; LFM3-VL-2B's average score across 62 vision benchmarks rises from 57.6 to 71.4. All models are available for download on Hugging Face, along with a single-GPU fine-tuning notebook.

OpenAI Releases GPT-6.1 Terra, Standard API Price Cut 50% Versus Previous Generation

OpenAI has released a new model, GPT-6.1 Terra, with standard API input pricing reduced to $1 per million tokens and output at $5, a 50% decrease compared with GPT-5.6 Terra. The model focuses on visual understanding and computer operation, scoring 50.3% at the highest reasoning effort on the Terminal-Bench 4.0 coding benchmark, up from 21.5% for the previous generation. OpenAI says the most complex coding tasks are still better suited to GPT-6.1 Sol or GPT-6 Astra.

On the same day Anthropic filed for its IPO, OpenAI released GPT-6.1 Terra and...

On October 26 local time, on the same day Anthropic submitted its IPO registration filing to the SEC, OpenAI released its new model GPT-6.1 Terra, with standard API prices cut by 50% compared with the previous generation. The model focuses on visual understanding and computer operation, with standard API pricing at $1 per million input tokens, $5 for output, and $0.05 for cached input; on the Terminal-Bench 4.0 coding benchmark, it scored 50.3% at the highest reasoning effort, compared with 21.5% for the previous generation. OpenAI said Terra has approached flagship-model levels on some tasks, but the most complex coding tasks are still better suited to GPT-6.1 Sol or GPT-6 Astra.

OpenAI releases GPT-6.1 Terra, focused on low-cost computer operation capabilities

OpenAI has released GPT-6.1 Terra, with API pricing at $1 per million input tokens and $5 per million output tokens, down 50% and about 58% respectively from the previous generation, and half the price of GPT-6.1 Sol. OpenAI says its task completion rate on Terminal-Bench 4.0 rose from 21.5% to 50.3%, and its error rate at the highest reasoning level on difficult factual tests dropped to 4.3%. Third-party organization Artificial Analysis measured an intelligence index of 48 points, lower than Sol's 52 points and Claude Opus 5.5's 58 points, but the average cost per task fell from $1.40 to $0.40. The model is now available via API, and Plus, Pro, Enterprise, and Edu users can use it in ChatGPT Work and Codex; ordinary Chat conversations are not yet supported.

OpenAI Releases GPT-6.1 Terra, API Price Cut to Half of GPT-5.6 Terra

OpenAI has released GPT-6.1 Terra, with API pricing at $1 per million input tokens and $5 per million output tokens, both half those of GPT-5.6 Terra, while cached input drops to $0.05 per million tokens. The model is now available in ChatGPT Work, Codex, and the API to Plus, Pro, Enterprise, and Edu users, focusing on visual reasoning, computer control, and writing. Artificial Analysis data shows that at its highest reasoning tier it scores 6 points higher than GPT-5.6 Terra on the Intelligence Index, with a per-task cost of $0.40 versus $1.40; its Terminal-Bench 4.0 score rose from 21.5% to 50.3%.

Moonshot AI Releases Kimi K3.1 Pro: 3 Trillion Parameters, 1 Million Context, Coding Agent Benchmarks Surpass GPT-6 Astra

Moonshot AI officially launched K3.1 Pro, the first model in the Kimi K3.1 series. It is a 3-trillion-parameter model supporting a 1-million-token context, and a Highspeed version with roughly 2–3x the output speed is being released alongside it. On Terminal-Bench 4.0, K3.1 Pro scored 65.7%, surpassing GPT-6 Astra's 58.2% and coming close to Claude Opus 5.5's 66.4%; on OfficeQA Pro v2 it scored 64.4%, achieving the current best. Starting today, it is available via kimi.com, the Kimi App, Kimi Work, Kimi Code, and the Kimi API. The full model weights will be released by November 4. API input pricing is 1.3 yuan per million tokens (cache hit) and 13 yuan (cache miss), with output at 68 yuan.

Moonshot AI releases Kimi K3.1 Pro: 2.8T total parameters, 104B activated parameters, weights scheduled to be opened on November 4

Moonshot AI releases Kimi K3.1 Pro, a multimodal MoE model with 2.8T total parameters, 104B activated parameters, and a 1 million token context window. The Standard and Highspeed versions are now available on the Kimi website, App, Kimi Work, Kimi Code and API. The company says Highspeed offers 2–3x faster output at twice the price of the Standard version. Artificial Analysis ranks it fourth on the Intelligence Index with a score of 55, behind Opus 5.5's 58; on Arena's Code Arena WebDev it ranks 8th with 1692 points, up nine places from K3. Standard pricing is $2 per million tokens for input, $10 for output, and $0.2 for cached input, one-third lower than K3; the weights have not yet been released, and the company promises to open them on November 4.

Moonshot AI Releases Kimi K3.1 Pro, Scoring 65.7% on Terminal-Bench 4.0

Moonshot AI has released Kimi K3.1 Pro, its first updated version, which scores 65.7% on Terminal-Bench 4.0—0.7 points lower than Claude Opus 5.5 and 7.5 points higher than GPT-6 Astra—while its average cost per task is lower than both. The model is a 3T-class open-weight MoE model with 2.8T total parameters and 104B active parameters, supports a 1M-token context and multimodality; in a tool-augmented hallucination stress test, its false-claim rate dropped from 23.4% in K3 to 8.3%. K3.1 Pro and the 2–3× faster Highspeed version are now available in the Kimi app and API, with full weights scheduled for release on November 4.

Moonshot AI releases Kimi K3.1 Pro: a 3T-class open-weight model with a 1 million token context

Moonshot AI releases Kimi K3.1 Pro, the first model in the K3.1 series and its strongest 3T-class open-weight model to date, with a 1 million token context window, alongside Kimi K3.1 Pro Highspeed, which is 2-3 times faster. The model is now available on Kimi.com, Kimi Work, Kimi Code and the Kimi API, with full model weights to be released on November 4. On Terminal-Bench 4.0, K3.1 Pro scores 65.7%, higher than GPT-6 Astra's 58.2% and slightly below Claude Opus 5.5's 66.4%; on the internal benchmark Kimi Code Bench 3.0 Long-horizon it scores 61.2%, higher than K3's 55.95%.

MiniMax open-sources the M3.1 Flash model, reducing usage cost by 70%

MiniMax announced on October 23 the open-sourcing of the M3.1 Flash model, which processed over 100 trillion tokens within 72 hours of launch. Compared with the previous generation M3, M3.1 Flash has comprehensively improved capabilities, with usage cost reduced by 70% and online inference speed increased by nearly 3 times. MiniMax stated it will work with chip manufacturers such as Huawei Ascend, NVIDIA, and AMD, as well as inference frameworks and communities such as vLLM, SGLang, and Hugging Face, to advance day-one open-source integration.

MemTensor Open-Sources Native Memory Model MemT-1.0, Available in 36B and 288B Tiers

The MemTensor team has open-sourced MemT-1.0, a model built on a native memory architecture, offering two size tiers—36B Nano and 288B Flash—and has released a 66-page technical report. The model connects a Native Memory Block in parallel alongside the standard Transformer Block, uses an erase gate, write gate, and forget gate to achieve token-level dynamic state updates, and extends memory capacity through a Mixture-of-Storage mechanism. According to the figures in the article, under a 256K history, VRAM drops from 5.06 GiB to 40 MiB, and first-token latency remains at 104–112 ms between 512 and 256K.

Baseten Base Labs releases decision models mammoth-1 and elephant-1

Baseten's Base Labs releases two decision models, elephant-1 (9B) and mammoth-1 (27B), both fully fine-tuned from Qwen3.5, with model weights publicly released. mammoth-1 ranks first among 208 models on the Jev Decision Index 0.4 with a score of 71.24, leading the second-place Perplexity Decider v1.2 by 1.8 points; elephant-1 scores 62.65 on that leaderboard and ties for first with d1-ultra on the Artificial Analysis Decision Index with a score of 65. Training used the eighth version of the data mixture with 6.21M samples, and replaced the full-attention layers with a block bidirectional attention mask; under fixed data, this change raised the public 10k-sample score from 44.8 to 48.6.

English translation is being prepared.

English translation is being prepared.

NVIDIA Releases Nemotron 3.5 Series of Open Models, Covering Embedding, Reranking, and Document Parsing

On October 21, 2026, NVIDIA expanded the Nemotron Retriever series, launching Nemotron 3.5 Embed (8B BF16, 1B BF16, 1B NVFP4), Nemotron 3.5 Rerank 8B BF16, and the 631-million-parameter Nemotron 3.5 Parse, aimed at multimodal retrieval and document processing for enterprise agent workflows. NVIDIA says that adding Rerank to the agent retrieval pipeline reduces the token cost per query by about 14% (with the 8B embedding model) and 13% (with the 1B model), respectively, while also improving retrieval accuracy. Parse scores 75.20 on ParseBench and processes 10.5 pages per second on a single H100; the open weights, data, and training recipes have been released on Hugging Face and can be deployed via NVIDIA NIM as well as Baseten, DeepInfra, and FriendliAI.

English translation is being prepared.

Memory Tensor Open-Sources MemT 1.0 Series Memory-Native Models

Memory Tensor has open-sourced and released the MemT 1.0 series of memory-native models, including MemT-1.0-Nano with 36B parameters and 3B activated (based on Qwen3.6-35B-A3B) and MemT-1.0-Flash with 288B parameters and 17B activated (based on DeepSeek-V4-Flash), adding a memory computation pathway parallel to attention in selected Transformer modules. The team built a 34-billion-token memory training corpus, of which 20 billion tokens were used for continued pre-training and 14 billion for post-training. In a comprehensive evaluation across 11 memory benchmarks, Flash and Nano achieved average scores of 50.6 and 47.0, respectively, ranking in the top two among the parametric memory methods evaluated; on 500 SWE-bench Verified tasks, the proportion of tasks that Nano succeeded on all three attempts rose from 54.00% for the no-experience baseline to 62.80%.