NAll news

All news

Tuesday46 items

  1. heise onlineModel releases
    AI score80

    Moonshot AI releases Kimi K3.1 Pro, AA Intelligence Index rises to 55 points, ranking fourth

    Moonshot AI has released Kimi K3.1 Pro, which rose from 44 points for Kimi K3 to 55 points on the Artificial Analysis Intelligence Index, ranking fourth and trailing Claude Opus 5.5's 58 points. The model retains the MoE architecture with 2.8 trillion parameters and 104 billion activated parameters, has a context window of 1 million tokens, and supports text, image, and video input. The success rate on Terminal-Bench 4.0 rose from 13% to 67%, API input and output prices are $2 and $10 per million tokens respectively, and the weights are scheduled for release on November 4.

    Same story · 6 reports

    #Model releases#Open-source ecosystem#Agents

Monday13 items

  1. The Rundown AIModel releases
    AI score75

    Moonshot Releases Kimi K3.1 Pro, Ranking Fourth on the Artificial Analysis Intelligence Index

    Moonshot has launched Kimi K3.1 Pro, an upgraded version of July's K3. The company says it matches the strongest closed-source models on key benchmarks, and freely downloadable weights will be released by November 4. Artificial Analysis ranks it fourth on the Intelligence Index, ahead of OpenAI's GPT-6 Astra and behind only Anthropic's two 5.5-series models and Meta's Muse Wonder. Pricing is $2 per million input tokens and $10 per million output tokens, comparable to Sonnet 5.5, but its token consumption rate is higher than that of frontier rivals; AA also rated it the best publicly available cyber-defense model, with a score of 64% and no safety refusals.

    Same story · 34 reports

    #Model releases#Open-source ecosystem#DeepSeek

  2. gihyo.jpModel releases
    AI score87

    Moonshot AI releases the open-weight MoE model Kimi K3.1 Pro, and launches a 2 to 3 times faster…

    On October 23, 2026, Moonshot AI released the multimodal MoE model Kimi K3.1 Pro, with a context length of 1 million tokens, 2.8T total parameters, and 104B activated parameters, while also launching Kimi K3.1 Pro Highspeed with a 2 to 3 times speed improvement. This version restructured the reinforcement learning process, surpassing GPT-6 Astra on Terminal-Bench 4.0 and slightly below Claude Opus 5.5, tying with GPT-6 Astra and Claude Opus 5.5 for SOTA on OfficeQA Pro v2, and reducing hallucinations in financial scenarios by 65%. Both models are now available on Kimi.ai, Kimi Work, Kimi Code, and the API, with open-source weights planned to be publicly released by November 4, 2026.

    Same story · 34 reports

    #Model releases#Open-source ecosystem#Multimodal

  3. Thomas WolfModel releases
    AI score63

    BOSS Zhipin research team Nanbeige releases open-source 33B model Nanbeige4.3-33B

    Nanbeige, a research team under BOSS Zhipin, has released the open-weight model Nanbeige4.3-33B, outperforming Qwen 27B on most agent benchmarks. The model achieves SOTA agentic performance among open-source models at roughly the 30B parameter scale, shows strong results on agentic coding and collaboration benchmarks, and after integration with mainstream agent frameworks can handle real-world tasks at lower inference cost. It will also serve as a teacher model for the subsequent smaller Nanbeige4.3-3B.

    Quoted postNanbeige · @nanbeige

    🚀Excited to announce Nanbeige4.3-33B! To push the performance ceiling of small agents, our strategy in Nanbeige4.3 starts with building a powerful teacher first. Nanbeige4.3-33B delivers SOTA agentic performance among open models at ~30B parameters: * Strong results on agentic coding and cowork benchmarks. * When integrated into mainstream agent harnesses, it tackles real-world tasks with cost-efficient inference. * It will serve as a teacher for upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B. https://t.co/tSToQgZf2O https://t.co/D228euvybz #LLM #Agent

    Same story · 3 reports

    #Model releases#Open-source ecosystem#Agents

  4. QbitAIModel releases
    AI score58

    Yanxi Dynamics Releases Embodied Foundation Model Paimon-0, Tops Four Leaderboards Including RoboColiseum

    Shenzhen-based company Yanxi Dynamics (Primotion), founded three months ago, has released its embodied foundation model Paimon-0, ranking first on RoboColiseum with a score of 0.7754 and scoring 0.6613 on the general manipulation sub-leaderboard, and topping RoboCasa365 with 58.1 points and VLA-Arena with an average success rate of 0.73. The model is centered on a World Action Model (WAM) and uses query token distillation to compress the inference time of a single action chunk from about 3000ms to 300ms. The company positions itself in embodied chips, proposes a device-cloud division of labor with a local brain plus a meta-brain, and plans to begin delivering its self-developed chips in 2028.

    #Model releases#Embodied AI#Evaluations / benchmarks

  5. QbitAIModel releases
    AI score79

    RockAI Technology, together with HKUST and CMU, open-sources EdgeRock-4B, reducing KV Cache by 99.1%

    Yanxin Matrix, under RockAI Technology, together with universities including HKUST and CMU, has released and open-sourced EdgeRock-4B. At a 256K context, its KV Cache is fixed at 72MiB, a 99.1% reduction compared with the source model Qwen3.5-4B. A single 80GB GPU can run 512 concurrent long-inference streams, with an output throughput of 12,124 tokens/s, 6.42 times that of the source model; the average pass@1 on AIME 2024–2026 is 82.9%. The model is derived from Qwen3.5-4B through staged conversion, requiring no pretraining from scratch; the weights and inference code have been open-sourced and are available for commercial use, supporting text and image inference.

    #Model releases#Open-source ecosystem#Reasoning

  6. 36Kr · AIModel releases
    AI score73

    Inception Labs releases diffusion LLM Mercury 3, up to 8x faster than GPT-6 Luna

    Inception Labs releases the diffusion LLM Mercury 3. In official benchmarks, it scores 89.23 on GPQA Diamond, higher than GPT-6 Luna's 86.36 and Claude Haiku 5.5's 84.51. Its output speed reaches 1,068 tokens per second, up to about 8x that of GPT-6 Luna, its context window has been expanded to 1 million tokens, and during the launch period it costs $0.10 per million input tokens and $0.35 per million output tokens. Its total parameter count is more than 40x larger than Mercury 1, its overall capability is about 3x higher, and its speed still remains at the thousand-tokens-per-second level.

    #Model releases#Reasoning#Agents

  7. TMTPost · AGIModel releases
    AI score66

    Kimi K3.1 Pro Released, Ranks Fourth Globally on Artificial Analysis Intelligence Index

    Moonshot AI released Kimi K3.1 Pro in the early hours of October 24. Artificial Analysis ranks it fourth globally on its intelligence index and judges that once the weights are released it will top the open-weight model leaderboard. On Terminal-Bench4.0, K3.1 Pro scores 65.7, second only to Claude Opus5.5's 66.4. The article also notes that K3.1 Pro completed its upgrade amid intensified anti-distillation efforts by Anthropic and other vendors, which the author views as a counterexample to the claim that Chinese models rely entirely on distillation.

    Same story · 6 reports

    #Model releases#Open-source ecosystem#Evaluations / benchmarks

  8. English text is being prepared.Model releases
    AI score52

    China's AI 3D model Hi3D V1.0 goes open source, releasing weights and inference code

    China's AI 3D creation platform Hi3D has announced the open-sourcing of the V1.0 model weights and inference code, with the resources now available on GitHub and Hugging Face. The model is based on the Sparc3D architecture, and its training data has been expanded from about 500,000 to about 2 million curated 3D assets. It supports an effective sparse grid resolution of up to 1536³ and takes a reference image and a coarse mesh as input to progressively refine the geometric structure, but it outputs only the geometric model and does not generate textures, materials, or vertex colors. V1.0 uses a custom license agreement: upon exceeding 200,000 monthly active users or reaching US$1 million in monthly revenue, the licensee must notify the licensor within 60 days and apply for a separate written authorization.

    #Model releases#Open-source ecosystem#Hugging Face

  9. English text is being prepared.Model releases
    AI score71

    Moonshot AI Releases Kimi K3.1 Pro, Strengthening 3D Modeling and Spatial Reasoning

    Moonshot AI has released Kimi K3.1 Pro, strengthening 3D creation and spatial reasoning capabilities, and launched the 3D Make feature, which supports online preview and editing of 3D models, downloading files for 3D printing, or exporting as videos and Live Photos. The model is the first product in the K3.1 series, supports a 1 million Token context, and can write and run scripts for software such as Blender, SolidWorks, and Rhino to complete modeling. Official BenchCAD results show that, in a test subset consisting of 1,000 files, the voxel IoU after tool invocation was 92.7%, versus 72.8% without tools; the model is currently integrated into the Kimi web client, Kimi Work, Kimi Code, and the API, the Highspeed version outputs at roughly 2 to 3 times the speed of the standard version, and the full model weights are planned for release before November 4.

    Same story · 34 reports

    #Model releases#Multimodal#Reasoning

  10. CaixinModel releases
    AI score66

    Moonshot AI launches Kimi K3.1 Pro, a post-trained version of K3

    In the early hours of October 24, Moonshot AI (Kimi) officially launched Kimi K3.1 Pro. This is the post-trained version of Kimi K3 and its most capable 3-trillion-parameter model to date, supporting a 1 million context window. Kimi says the model shows particularly notable improvements in coding, 3D creation, financial research and other long-horizon Agent capabilities. This is the first time Kimi has introduced a Pro version in its model versions, and K3.1 Pro is also the first model in the K3.1 series, leaving room for subsequent releases of cost-effective models.

    Same story · 34 reports

    #Industry news#Model releases#Agents

  11. MarkTechPostModel releases
    AI score60

    Nanbeige Releases Nanbeige4.3-33B Open-Source Agentic Coding Model

    Nanbeige LLM Lab released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 license. It is a 33B dense model aimed at agentic coding and office collaboration tasks, supports a 524,288 token context window, and is released alongside Nanbeige4.3-33B-Base, which supports customizable post-training. In the lab's self-built comparison table, it leads in 17 of 20 benchmarks against Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B, scoring 80.9 on Terminal-Bench 2.1, but trails Qwen3.8-27B on SWE-Bench Pro with 60.3 versus 61.7. The article points out that some competitor figures in its table are inconsistent with the official model cards from Meta and Google, and that two SWE-Bench Pro entries appear to be swapped; these results need to await independent reproduction. The model is positioned as the teacher model for the Nanbeige4.3 series, with plans to subsequently distill agentic capabilities into Nanbeige4.3-3B.

    Same story · 3 reports

    #Model releases#Open-source ecosystem#Agents

  12. MarkTechPostModel releases
    AI score66

    Hugging Face Releases 14.8B Open-Source DNA Language Model Carbon-X

    Hugging Face Bio has released Carbon-X, a 14.8B-parameter decoder-only DNA language model that can directly read raw DNA and answer in English, with weights released under Apache 2.0. The model is based on Qwen3-14B and uses a hybrid tokenizer that adds 5,515 DNA tokens, supports a 40,960-token context, and can run in bfloat16 on a single 80GB GPU; the official GGUF Q8_0 is 15.8GB. Hugging Face self-reports an average of 66.7% across 5 DNA benchmarks, with BioReason Coding reaching 74.9%, but only 29.8% on GenomeQA splice-site questions, and it has not undergone clinical validation.

    Same story · 9 reports

    #Model releases#Open-source ecosystem#English text is being prepared.

  13. Testing CatalogModel releases
    AI score62

    Nanbeige Releases the Open-Source Nanbeige4.3-33B Model

    Nanbeige has released Nanbeige4.3-33B, claiming that it achieves SOTA agent performance among open-source models with roughly 30B parameters, with outstanding results on agentic coding and cowork tasks. The model surpasses Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B on most benchmarks, and once integrated into mainstream agent frameworks, it can handle real-world tasks at relatively low inference cost. The company also plans to release a smaller Nanbeige4.3-3B, with the 33B version serving as its teacher model.

    Quoted postNanbeige · @nanbeige

    🚀Excited to announce Nanbeige4.3-33B! To push the performance ceiling of small agents, our strategy in Nanbeige4.3 starts with building a powerful teacher first. Nanbeige4.3-33B delivers SOTA agentic performance among open models at ~30B parameters: * Strong results on agentic coding and cowork benchmarks. * When integrated into mainstream agent harnesses, it tackles real-world tasks with cost-efficient inference. * It will serve as a teacher for upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B. https://t.co/tSToQgZf2O https://t.co/D228euvybz #LLM #Agent

    Same story · 3 reports

    #Model releases#Open-source ecosystem#Agents

Sunday6 items

  1. The DecoderModel releases
    AI score60

    Block releases speech model Dessa, trained entirely on human speech

    Block has released the speech model Dessa, which it says is the first AI model trained from scratch entirely on human speech, without any text data. Dessa processes audio directly rather than first converting it to text, allowing it to capture emotion, sarcasm, laughter, and pauses in real time. It supports full-duplex communication and can run locally on a MacBook Pro M4 Max, with training performed on a single NVIDIA GB200 NVL72. The model has not yet been made public, and the team is recruiting AI safety researchers to test potential risks.

    Same story · 10 reports

    #Model releases#Speech#English text is being prepared.

  2. IT Home · Artificial IntelligenceModel releases
    AI score62

    Block launches full-duplex speech model Dessa, supporting local operation on MacBook

    Block launches the full-duplex speech model Dessa, which can listen to the user's speech while speaking; emotional expression is natively built into the model, and it can simulate states such as laughter, pauses, and choking up. The company says the model can run locally on consumer-grade computers, with the device threshold being MacBooks with M4 Max or higher configurations, and that training used only a single rack of NVIDIA GPUs. At this stage Dessa is not open for public trial; Block is sending testing invitations to safety researchers, focusing on risk scenarios in which such speech agents deceive humans.

    Same story · 10 reports

    #Model releases#Speech#English text is being prepared.

  3. Shao MengModel releases
    AI score76

    Block releases Dessa, a full-duplex model trained purely on speech, which can run locally in real time on a MacBook Pro M4 Max

    Block, founded by Twitter co-founder Jack, has released Dessa, a speech-to-speech model trained from scratch purely on human speech, with no text and no pretrained base, supporting full duplex so it can listen and speak at the same time. The model claims inference faster than real time and can run locally on a MacBook Pro M4 Max and above, and the entire model is said to have been trained on a single NVIDIA GB200 NVL72. The team says its goal is to let the model understand tone and emotion, avoiding the safety risk of continuing to respond without perceiving the other person's distress.

    Quoted postjack · @jack

    We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. https://t.co/r0qFLdvd6c

    Same story · 10 reports

    #Model releases#Speech#English text is being prepared.

  4. Phoenix TechnologyModel releases
    AI score66

    Block launches full-duplex speech model Dessa, supports local operation

    Block has launched the full-duplex speech model Dessa, which can listen to the user's speech while speaking, and can simulate emotional expressions such as laughter, pauses, and choking up. The company says it is trained entirely on human speech data, using only a single rack of NVIDIA GPUs for training, and can run locally on MacBooks with M4 Max and higher configurations. At this stage Dessa is not open for public trial; Block is sending testing invitations to security researchers, focusing on risk scenarios in which speech agents deceive humans.

    Same story · 10 reports

    #Model releases#Speech#English text is being prepared.

  5. Hacker News TrendingModel releases
    AI score80

    Dream Labs Releases Robot Model Dream-1 and Opens Public API

    Dream Labs has released the world-action model Dream-1, which can control unseen dual-arm parallel-jaw gripper robots to perform unseen tasks via a public API, without fine-tuning or demonstration data from that robot. The company also announced it is emerging from stealth with a $90 million seed round. Dream-1 costs $3 per hour to run, and Dream-1-flash costs $2 per hour; each call returns 1.6 seconds of actions. If zero-shot performance is insufficient, it can be fine-tuned via FaaS with 5 minutes of robot data.

    Same story · 2 reports

    #Model releases#Embodied AI#Open-source ecosystem

  6. TechCrunch AIModel releases
    AI score67

    Block releases full-duplex voice model Dessa, which can run locally on consumer-grade GPUs

    Block's research lab has released the voice model Dessa, which uses a full-duplex architecture, can listen and speak at the same time, and can run locally on a MacBook M4 Max or a more powerful consumer-grade computer. The model was trained entirely on human speech and a single NVIDIA GPU, and the company says expressions such as emotion, pauses, and laughter are built into the model rather than driven by a text pipeline. Dessa is not currently publicly available, and Block is recruiting security researchers to test the risk that it could be used to deceive conversation partners.

    Same story · 10 reports

    #Model releases#Speech#English text is being prepared.