AI News

AI changes every day,
catch up on it all on one page

The latest updates from AI companies, research institutions, and media around the world, automatically selected for you and updated around the clock.

Updated 11/17 00:43 (UTC+8) · refreshed hourly

10/20 · Tuesday5 items

Artificial Analysis@ArtificialAnlysModelsAI Score66

DeepSeek V4.1 Flash 1017 released, scoring 45 on the Artificial Analysis Intelligence Index and tying for second among open-source models

DeepSeek released DeepSeek V4.1 Flash 1017, which scored 45 on the Artificial Analysis Intelligence Index, 6 points higher than DeepSeek V4.1 Flash. It tied with GLM-5.3 for second place among open-weight models, behind only MiMo-V2.6-Pro.

Why it's recommended:An independent evaluator provides specific scores across multiple benchmarks and cost comparisons, allowing readers to assess the model's value among open-source alternatives.

Arena.ai@arenaModelsAI Score65

Meta Muse Wonder Full (Max) ranks sixth on Code Arena: WebDev with 1,726 points.

Arena announced that Meta's Muse Wonder Full (Max) entered the Code Arena: WebDev leaderboard for the first time, ranking sixth with 1726 points and making Meta the third laboratory to hold this position; Google previously held it.

Why it's recommended:The ranking organization published Meta's new model's debut results and category rankings on Code Arena WebDev, allowing readers to compare the current standings of models from different labs.

10/19 · Monday3 items

Ant Group inclusionAI: New GitHub RepositoryModelsAI Score72

Ant's inclusionAI open-sources the Ming-3.0-tts speech synthesis model

Ant Group's inclusionAI releases Ming-3.0-tts on GitHub, including two speech-generation models: a 0.5B dense model and a Ling V3-based tiny model with 7.9B total parameters and 1.3B active parameters. They are open-sourced under the MIT license, with weights available for download from Hugging Face and ModelScope.

Why it's recommended:The official repository provides the leaderboard rankings, complete task lists, and deployment details for both model versions, allowing readers to assess their practical usability for multilingual and paralinguistic control.

Also reported by 3 other sources
MiniMax:BlogModelsAI Score70

MiniMax released M3.1 Flash, nearly tripling inference speed and reducing usage costs by 70%

MiniMax releases MiniMax M3.1 Flash. The anonymous model Space Bunny was an early version and had processed more than 200 trillion tokens.

Why it's recommended:The official release provides specific benchmark scores, inference speeds, and pricing per million tokens, allowing readers to directly compare changes in cost-effectiveness for agent tasks.

Reported by 6 additional sources
Arena.ai@arenaModelsAI Score69

DeepSeek-V4.1-Flash-20261017 Tops the Open-Source Model Rankings in Code Arena: WebDev

Arena announced that DeepSeek-V4.1-Flash-20261017 (Max) became the top open-source model on the Code Arena: WebDev leaderboard, surpassing Kimi K3 (Max) at approximately one-tenth the cost.

Why it's recommended:The Arena leaderboard provides specific rankings, scores, and price points, allowing readers to evaluate the cost-performance position of DeepSeek's new model for WebDev coding.

Also reported by 1 other source

10/17 · Saturday2 items

DeepSeek: API ChangelogModelsAI Score62

DeepSeek-V4.1-Flash-1017 released, with improved scores across multiple benchmarks and unchanged API pricing

DeepSeek has upgraded V4.1 Flash to DeepSeek-V4.1-Flash-1017. The web version and app have been updated simultaneously, while the API access method and pricing remain unchanged; it is still accessed using deepseek-flash.

Why it's recommended:The original text lists the scores for each benchmark and explains that API access and pricing remain unchanged, allowing readers to directly assess whether the new model is worth switching to.

Reported by 6 additional sources
Arena.ai@arenaModelsAI Score80

Muse Spark Open (Max) Tops U.S. Open-Source Models on Code Arena: WebDev

Arena announced that Meta's Muse Spark Open (Max) scored 1511 on Code Arena: WebDev, ranking first among U.S. open-source models and 13th among open-source models overall.

Why it's recommended:Using Arena's own Code Arena leaderboard, the article provides specific scores, rankings, and pricing comparisons for Muse Spark Open, helping readers assess the open-source model's actual capabilities.

10/16 · Friday8 items

Full Artificial Analysis ArticleModelsAI Score61

Reflection releases its first open-weight model, Beam; Artificial Analysis gives it a score of 30, ranking it second among U.S. open-weight models

Artificial Analysis evaluated Reflection’s first model, Beam. It scored 30 on the Artificial Analysis Intelligence Index, ranking second among U.S. open-weight models, behind only Meta’s Muse Spark Open (39).

Why it's recommended:The third-party evaluation provides Beam’s ranking among U.S. open-weight models and token-efficiency data, which can be used to compare its agentic and coding capabilities side by side.

Also reported by 1 other source
Reflection BlogModelsAI Score83

Reflection releases its first open-weight model, Beam, a 501B-parameter MoE under the Apache 2.0 license

Reflection releases Beam, its first open-weight model, with weights available on Hugging Face under Apache 2.0. Beam is a sparse MoE model with 501B total parameters and 23B active parameters, designed for coding, reasoning, and long-horizon agent tasks. It offers weights in BF16, FP8, and NVFP4 precision, along with a drafter model for speculative decoding.

Why it's recommended:The original article provides specific information on weights, precision formats, ecosystem support, and self-hosting options, which can be used to evaluate implementation and model selection.

Also reported by 2 other sources
Full Artificial Analysis ArticleModelsAI Score66

Artificial Analysis releases initial AA-AgentPerf agent inference benchmarks for NVIDIA Vera Rubin NVL72

Artificial Analysis published its initial AA-AgentPerf results, testing NVIDIA Vera Rubin NVL72's serving efficiency on agentic coding workloads against GB300 NVL72.

Why it's recommended:Artificial Analysis measured the throughput per watt of four models on Vera Rubin NVL72 and GB300, allowing readers to assess the impact of new hardware on the cost of agent inference services.

Also reported by 1 other source
WeChat Official Account: Digital Life KazikeModelsAI Score74

Kling 4.0 launched late at night, supporting direct generation of 30-second single-shot videos

Kling AI officially launches its Kling 4.0 video model, following early access to 4.0 Flash for selected Black Gold annual pass and Ultra annual subscribers in late September. After hands-on testing, the author identifies three main improvements: direct generation of single-take videos up to 30 seconds long, with better details such as physical deformation when characters interact with cardboard boxes in complex scenes like urban chases and circuses; noticeably improved action continuity and camera immersion over the previous generation in large-scale motion and complex fights; and a shift from first-and-last-frame guidance to multi-frame references with up to 10 frames, while All-in-One Reference can replace camera movements, scenes, and characters.

Why it's recommended:The author spent credits testing multiple types of scenarios and provided credit pricing and generation costs, allowing readers to assess its usability in real-world creative workflows.

Arena.ai@arenaModelsAI Score78

Muse Spark Open is now available on Arena and can be tested in Agent Arena; ratings coming soon

Arena announced that Meta's Muse Spark Open has entered Agent Arena and is available for user testing, with scores to be announced soon. The model has 816B total parameters, 42B active parameters, a 1M-token context window, and native multimodal capabilities.

Why it's recommended:Arena officially announced that Meta's new model has entered Agent Arena and Code Arena, allowing readers to test it themselves and await scores based on real tasks.

PyTorch:BlogModelsAI Score83

Meta releases Muse Spark Open, an open-weight model with 816B parameters and a 1M context window

Meta released the open-weight model Muse Spark Open, now available on Hugging Face under the Muse Model License. The model has 816B total parameters and 42B activated parameters, a 1M-token context window, and is designed for coding and agent workflows. It is natively multimodal, supporting images, audio, and documents, and adds safety post-training compared with Muse Spark, released in August.

Why it's recommended:Meta has released the weights for Muse Spark and provided Helion kernel and TorchTitan fine-tuning support, allowing readers to assess its deployment and customization paths.

Reported by 14 additional sources
ARC Prize@arcprizeModelsAI Score66

Gemini 4 Argon results released for the ARC-AGI (Verified) evaluation: 97.5% on ARC-AGI-3

ARC Prize released the evaluation results for Gemini 4 Argon on ARC-AGI (Verified). On ARC-AGI-3, it scored 97.5% with the provider adapter harness at a cost of $4.8K, and 49.8% with the standard harness at a cost of $5.8K; on ARC-AGI-2, it scored 91.3% at $0.32/task; and on ARC-AGI-1, it scored 98.5% at $0.16/task. Google DeepMind is the second frontier AI lab to score above 95% on ARC-AGI-3.

Why it's recommended:ARC Prize released Gemini 4 Argon's scores and costs across the three versions of ARC-AGI, allowing readers to compare the differences between various harness configurations.

Also reported by 2 other sources
Microsoft AI: Official BlogModelsAI Score66

Microsoft AI reveals pretraining details for MAI-Base-2, a 2.8T-parameter MoE model with 4.2 to 12.3 times the FLOPs efficiency of DeepSeek-V4-Pro-Base

Microsoft AI announced that it has completed pretraining MAI-Base-2, an MoE foundation model with 80 billion active parameters and 2.8T total parameters. On held-out evaluations, its BPB was lower than that of all open-weight models tested, and its pretraining FLOPs efficiency was 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base.

Why it's recommended:The official post details MAI-Base-2's complete pretraining methodology, offering engineering insights readers can draw on, including the scaling ladder, Muon optimization, and multi-round dropless MoE.

Also reported by 2 other sources

10/15 · Thursday10 items

StepFun@StepFun_aiModelsAI Score67

Step 5 Preview launched on FeatherlessAI, with the official account reposting information about open-weight model deployment

StepFun shared a post from FeatherlessAI announcing that Step 5 Preview is now available on the platform, served at FP8 precision. The open-weight model has a 600B sparse MoE architecture, activating 27B parameters per token, and is designed for long-horizon agent work. It delivers frontier-level performance in software engineering and professional knowledge work, especially finance. According to Artificial Analysis data, its cost per task is about 1/6 that of Claude Opus 5.5, with intelligence comparable to Kimi K3.

Why it's recommended:The cited material provides Step 5 Preview's parameter count, cost comparisons, and deployment access points, allowing readers to assess its value for money and how to access it.

Black Forest Labs:BlogModelsAI Score77

Black Forest Labs released the open-source text-to-image model FLUX 3 Image Klein (5.6B)

Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weight model supporting text-to-image generation, image editing, and multi-reference generation with up to 10 reference images. It uses the FLUX Kommunity License.

Why it's recommended:The official post reports the model's parameter scale, memory requirements, generation speed, and evaluation comparisons, helping readers assess its suitability for local deployment and fine-tuning.

Also reported by 3 other sources
NVIDIA BlogModelsAI Score71

NVIDIA releases open-source Nemotron 3.5 Super VL and two Bolt edge models

NVIDIA released Nemotron 3.5 Super VL, an open-source multimodal model with 120B parameters designed for high-frequency agent tasks. NVIDIA says its accuracy is comparable to that of its largest reasoning model, with roughly one-fifth the parameters and up to 5 times the throughput of the next-best comparable model. It also introduced two edge models: Bolt 74B-A9B 24GB, with native 2-bit quantization that can run locally on RTX/DGX Spark, and Bolt E4B 4GB, launching on October 23.

Why it's recommended:The article outlines the deployment tiers, quantization, and training data scale targeted by each of the three open models, helping readers determine which tier suits their agent use case.

Reported by 8 other sources
StepFun@StepFun_aiModelsAI Score66

StepFun demonstrates the inference performance of Step 5 Preview on NVIDIA GB300

StepFun says Step 5 Preview delivers excellent inference performance on NVIDIA GB300 and looks forward to seeing the applications developers build with it. The NVIDIA post it shared describes the model as StepFun's largest open-weight release, featuring a 600B MoE architecture capable of handling agent tasks lasting up to 24 hours. It achieves nearly 8K tokens/sec/GPU on GB300 using NVFP4, offers an NVFP4 checkpoint (https://huggingface.co/stepfun-ai/Step-5-Preview-NVFP4), and supports SGLang, vLLM, post-training with NeMo Automodel, and one-click deployment with NVIDIA NIM.

Why it's recommended:The original article adds details on Step 5 Preview's inference performance on NVIDIA GB300 and a link to its NVFP4 weights, helping developers evaluate deployment options.

StepFun@StepFun_aiModelsAI Score77

StepStar opens the Step 5 Preview weights

StepFun announced that Step 5 Preview is available with open weights. BF16 and FP8 weights are now available on Hugging Face and ModelScope, along with a technical blog post.

Why it's recommended:The official announcement provides weight specifications and download links, allowing readers who need open-source MoE weights to assess whether to incorporate them into their workflows.

Also reported by 3 other sources
Load more

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.