Skip to content
Live
#Model releases
Nov 10Tue
  2. Epoch AI

    How is Inference Performance Index (IPI) Measured for Inference Capability

    Nvidia GB300 offers a 25-fold improvement in capability compared to H100 for intelligent agent inference tasks, far exceeding the 2.5-fold increase in 8-bit FLOP/s. The IPI index measures inference capability by comparing the number of requests a chip can handle simultaneously under various output speed requirements. GB300 is set as the baseline value of 1, and chips scoring 1 can serve twice as many users under the same conditions. IPI version 1 combines results from four representative models, using GB300 as the reference benchmark.

  3. Artificial Analysis Articles66

    Meta's latest model Muse Ember scores 36 on the AI Index, comparable to Claude Haiku 5.5 but with lower costs

    Muse Ember scores 36, comparable to Claude Haiku 5.5 but with a task cost of only 0.14 USD, two-thirds lower than the latter; its input/output token prices are the same as Claude Haiku 5.5 and GPT-6 Luna, but caching input costs are lower. It scored 68.1% on AutomationBench-AA, comparable to Muse Wonder, but has the lowest knowledge accuracy (21.1%)

    Same story · 4 outlets
  4. TechCrunch · AI72

    Meta releases lightweight model Muse Ember, highlighting agent execution and low cost.

    Meta Superintelligence Lab releases lightweight model Muse Ember, priced at 10 cents per million tokens, focusing on high-frequency agent tasks such as tool calling, coding subtasks, routing, extraction, and long document parsing.

    Same story · 2 outlets

    Why it matters: Meta uses a pricing of 10 cents per million input tokens and the Artificial Analysis comprehensive benchmark to place the lightweight model in the agent execution context, comparing it with Haiku and Luna.

Nov 9Mon
  2. QbitAI82

    Tencent Hunyuan Hy4 released: 770B total parameters, 49B activation parameters, first supports native vision and computer operation

    Tencent Mixuan Hy4 is officially released, with a total of 770B parameters and 49B activation parameters. It introduces native visual capabilities and computer operation support for the first time, while also upgrading command compliance, long text understanding, and long-range Agent capabilities. The model has been launched on products such as Tencent Cloud TokenHub, WorkBuddy/CodeBuddy, and Yuanbao. API access has been provided to platforms like OpenRouter, Hermes, and OpenCode, with weights to be open-sourced in the near future.

    Why it matters: The author uses multiple long-range tasks to test the autonomous execution and self-check capabilities of the open-source model 'Hunyuan Hy4', thereby judging the actual level of open-source models in Agent delivery.

  3. ITHome · AI52

    Shizuku AI announces the full-duplex model Shizuku-Duplex, laying the foundation for AI companions

    Shizuku AI releases the full-duplex model Shizuku-Duplex, supporting intelligent Japanese dialogue and initiating thinking processes without complete inquiries. The model performs best in natural Japanese dialogue and has response quality second only to GPT-Live, adopting a combination of front-end interaction and back-end asynchronous processing architecture, supporting three output methods: audio, text, and actions.

  4. QbitAI47

    Ji Yuan Law Release NeoHorse-1.5-Flash Preview, Agent model RSI enters the second round of iteration

    Ji Yuan Law Release Agent Native Model NeoHorse-1.5-Flash Preview, based on Qwen3.8-Flash training, 180B total parameters, 6B activation parameters, score improvement in Agents' Last Exam, Natural Language to Repository and other evaluations. This model iterates through recursive self-improvement (RSI) and combines the OpenSquilla platform to achieve differential diagnosis of multiple models and online learning, now open for limited internal testing.

  5. ITHome · AI65

    Microsoft files AI patent: Real-time live stream voice modification and lip synchronization

    Microsoft filed a patent application, using AI to almost instantly modify live broadcast content and synchronize mouth movements, capable of replacing verbal fillers, stuttering, offensive words, etc., supporting real-time translation, with a delay of 1 to 10 seconds, potential application scenarios include lectures, sports event broadcasts, etc., currently only a patent application, the technology has not yet been authorized.

  7. MIT News · AI38

    MIT develops FLEX platform to help robots understand natural language instructions

    The FLEX platform developed by researchers from MIT and New York University enables robots to infer deep meaning from natural language instructions, enhancing their motion planning capabilities. The platform uses GPT-4o to analyze instruction patterns, allowing robots to avoid electronic devices such as laptops and phones in virtual scenarios without explicit instructions. The entire process takes about five seconds, aiming to help robots understand the underlying intent behind human instructions.

  9. ITHome · AI60

    LG AI Research Institute announces the free opening of the K-EXAONE 2.5 model API to the public

    LG AI Research Institute has opened the K-EXAONE 2.5 model API for free, allowing any user to use it at no cost for development and research. This self-developed model enhances the capabilities of the intelligent agent through additional training and performs equally well as DeepSeek-V4.1-Flash and GLM-5.3 in benchmark tests such as AutomationBench and SWE-Bench Pro.

  10. Artificial Analysis Articles38

    Introducing the Artificial Analysis Decision Index v2

    Artificial Analysis Decision Index v2 introduces a new metric rewarding model correctness and calibration confidence, with a 45% benchmark based on a private dataset; Drex Max v1.1 (Nace AI) leads with 73 points, GPT-6 Luna (Max) ranks first among language models with 76 points, but with significantly higher costs and inference time.

Nov 8Sun
  1. MiniMax · Hugging Face new models

    MiniMax-H3-Image-PE provides hint extension functionality

    MiniMax-H3-Image-PE uses prompt expansion (PE) to convert short requests into detailed image descriptions and recommends aspect ratios. It supports two workflows: text-to-image (T2I) and image editing, using fine-tuned FP4 weights from Qwen3.8-27B. The online API can be accessed via platform.minimax.io or platform.minimax.cn, and the documentation provides support in both Chinese and English.

  3. MarkTechPost58

    Nous Research releases Arete-27B, reducing the promotional content of Qwen3.8-27B from 22% to 0.9%

    Nous Research has released Arete-27B, a 27B parameter model with open-weight quantization, which removes promotional content about China from Qwen3.8-27B through supervised fine-tuning. In 803 prompt evaluations, the promotional content ratio decreased from 22% to 0.9%, while retaining the agent skills of the Qwen class and being free from political bias.

  4. MarkTechPost65

    How to use llama.cpp to run NVIDIA Dory 4B: Step-by-step guide

    Use llama.cpp to run the NVIDIA Dory 1.0 4B model, which achieves inference enhancement through repeated intermediate layer calculations. Each additional loop increases computational load without increasing parameters. Build the specified branch of llama.cpp, download the 3.7 GB GGUF file, analyze the parameters and cache consumption per loop, test decoding speed and cache impact, and evaluate accuracy, answer length, and inference time changes with loop counts on GSM8K.

  6. The Decoder65

    The thinking machine Inkling 2 Small used 316 profanities in benchmark tests.

    Vals AI researchers analyzed the thinking logs of Inkling 2 Small over 2,655 runs on 13 benchmarks and found 50 instances using 'fuck' or its variants, totaling 316 times, with 241 of them coming from two out-of-control runs. In one task, the model generated a single response of 500,000 characters while calculating asteroid trajectories, repeating the same sentence and including 158 instances of profanity. ProgramBench was the most affected, with profanity appearing in 20 out of 200 runs.

Nov 7Sat
  2. ITHome · AI65

    Meta launches a dedicated version of existing models based on Muse Wonder training, with proactive memory and focus capabilities delegated

    Meta has launched a dedicated version of existing models trained on Muse Wonder, transferring proactive memory and proactive focus capabilities to the Muse and Meta AI dedicated model groups. Muse models have increased the proportion of correct use of key context in 90% of tasks to 76%, and the score of Meta AI models in Meta's 'Personal Super Intelligence Index' has increased from 38% to 47%. New models will be prioritized for subscription users, with subscription prices of $20 per month or $100 per month.

  3. ITHome · AI22

    Wccftech reports that the AMD Zen 6 sample scored, with a 10-core multi-core performance about 21.5% ahead of the previous 12-core generation.

    AMD Zen 6 10-core sample scored 3100 single-core and 18313 multi-core in Geekbench 7, 21.5% higher multi-core performance than the 12-core Ryzen AI 9 HX 370; the 8-core sample's multi-core score of 15075 matched the latter, while the 6-core sample's multi-core score of 10670 was 17% higher than the Ryzen AI 5 340.

  4. 36Kr · AI70

    Google released the Gemini Omni 1.2 Flash video generation model, increasing the duration of a single segment to 15 seconds.

    Google has launched the AI video generation model Gemini Omni 1.2 Flash, natively supporting the longest 15-second video, with a duration limit increased by 50% compared to the previous generation's 10 seconds. Users can set the generation duration from 3 seconds to 15 seconds in 1-second increments.

    Same story · 6 outlets
  5. The Decoder65

    Recursive Intelligence releases open-source model Ripple-65B Preview, which performs excellently in coding benchmarks

    Same story · 4 outlets
  6. O'Reilly Radar22

    AI Weekly Scan: The cost of progress

    Google has released the Gemini 4 Carbon for cybersecurity and multi-step tasks, French Mistral has announced the weights of its 100 billion parameter Large 4 model, and DeepMind has previewed the Genie 4 world model for simulated environments.

    Same story · 10 outlets
Nov 6Fri
  2. The Decoder73

    Thinking Machines has released the Inkling 2 Small, claiming it is the strongest Western open-source weight model.

    Thinking Machines releases second-generation model Inkling 2 Small, using MoE architecture, with total parameters of 594B and activation parameters of 17B, supporting text and image input and adjustable thinking depth.

    Same story · 3 outlets

    Why it matters: The original text provides specific comparisons of the new model in terms of intelligence index, agent tasks, and cost, allowing readers to judge the competitive position of open-source weight models accordingly.

  5. ITHome · AI58

    Pangram officially releases AI image detection model Pangram Image 1.0, with an accuracy rate of 99.96%

    The Pangram Image 1.0 model locates AI-generated content using heatmaps, achieving an accuracy of 99.96% on a test dataset of 1200 photos. This model can resist the effects of compression, scaling, and artificial modifications, with an accuracy of 96.4% still maintained at 256×256 resolution. The simultaneously released Pangram Image Flash 1.0 video detection model achieves an accuracy of 98.7% on the DF26 dataset, leading the second place by 29 percentage points.

  6. Hugging Face Daily Papers65

    BestRand: A Stable and Low-Cost Benchmark Compression Method

    BestRand confirms high-probability strong performance of random sampling in benchmark tests through analysis of error distribution, but exists rare large errors leading to instability. The method is validated in over 70 text, multimodal, and agent benchmark tests, reducing high error risk by 50.2% compared to competing methods, decreasing average estimated error by 16.4%, while requiring only 51.3% of the construction cost, and averaging a 21.8% reduction in OOD estimated error.

  7. MarkTechPost71

    Thinking Machines has released the Inkling 2 Small: a MoE model with 594B total parameters and 17B activations.

    Thinking Machines Lab released the first model of the second-generation Inkling, Inkling 2 Small, a sparse MoE with 594B total parameters and 17B activation parameters, supporting text and image input and variable inference strength, with a context of 1M tokens.

    Same story · 3 outlets
  8. ITHome · AI55

    webAI's Open Source TwIL-8B Inference Model: Focused on Mathematics and Programming, Outperforms Gemma 4 31B in Long Board Performance

    webAI Launches TwIL-8B Inference Model, Optimized for Math and Code Tasks Based on IBM Granite-4.2-8B, Scores Better than Gemma 4 31B in MATH-500 and AIME 2025 Tests, Best in LiveCodeBench Group, IFBench on Par with Gemma 4 12B. Model Inference Capability Reaches 88~93% of Frontier Models, Reduces AI Usage Costs but Has Limited Knowledge Base.

    Same story · 2 outlets
  10. ITHome · AI

    Ricursive Intelligence releases Ripple-65B Preview, stating it has the best agent programming performance in the US open weight model.

    Ricursive Intelligence releases Ripple-65B Preview model, claiming it has the best agent programming performance in the US open weight model.

    Same story · 4 outlets
  11. MarkTechPost72

    Google releases Gemini Omni 1.2 Flash, supporting 15-second videos, priced at $0.07 per second for 720p

    Google releases video generation and editing model Gemini Omni 1.2 Flash, natively supporting up to 15-second clips, with 720p video pricing at $0.07 per second, a reduction of about 30% compared to Omni 1.1 Flash

    Same story · 6 outlets