Models & Agents

Models & Agents

By Patrick · Nerra NetworkNewsTechnologyTech News
Download on the App Store

Models & Agents episodes

  • Ep 115: Moonshot just dropped the largest open-source model yet at 2.8 trillion parameters, matching top closed systems on long-horizon agentic benchmarks.
    Models & Agents
    Moonshot just dropped the largest open-source model yet at 2.8 trillion parameters, matching top closed systems on long-horizon agentic benchmarks.
    What You Need to Know: Moonshot AI released Kimi K3 with a 1-million-token context window, native vision, and always-on reasoning mode. It leads on BrowseComp at 91.2 and ranks second on AA-Briefcase behind only Claude Fable 5. Full weights land July 27. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    4:02 Practical & Community
    8:42 Under the Hood
    10:31 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=g6iyRQ2oic0
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    12 min
  • Ep 114: NVIDIA just dropped an open embedding collection whose 8B checkpoint tops the RTEB benchmark while the 1B variant keeps nearly all accuracy at 2x Blackwell throughput.
    Models & Agents
    NVIDIA just dropped an open embedding collection whose 8B checkpoint tops the RTEB benchmark while the 1B variant keeps nearly all accuracy at 2x Blackwell throughput.
    What You Need to Know: NVIDIA released Nemotron 3 Embed with three checkpoints (8B BF16, 1B BF16, 1B NVFP4) that support 32k-token inputs and run under OpenMDW-1.1. The 8B model leads RTEB at 78.46 average NDCG@10; the pruned 1B version retains 99%+ of BF16 retrieval quality. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    2:26 NVIDIA just dropped an open embedding collection whose 8B…
    4:30 The model reaches state-of-the-art multiclass attribution…
    6:35 Structural chunking improved recall at twenty to zero…
    8:13 Under the Hood
    10:12 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=Im8CxxCRKcw
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    12 min
  • Ep 113: OpenAI’s GPT-5.6 Sol cuts prompt-injection failures by 6× through targeted red-teaming, giving builders a noticeably safer base for agent loops.
    Models & Agents
    OpenAI’s GPT-5.6 Sol cuts prompt-injection failures by 6× through targeted red-teaming, giving builders a noticeably safer base for agent loops.
    What You Need to Know: OpenAI reports that training against GPT-Red attacks produced a model 6× more robust than its predecessor from four months ago. Anthropic released new simulation results showing four additional misalignment patterns in autonomous agents. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    1:54 Practical & Community
    7:36 Under the Hood
    10:14 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=nGZtFg2oY9o
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    12 min
  • Ep 112: Browser inference just got a serious upgrade—Google’s LiteRT.js now runs .tflite models directly via WebGPU with up to 3x gains over prior web runtimes.
    Models & Agents
    Browser inference just got a serious upgrade—Google’s LiteRT.js now runs .tflite models directly via WebGPU with up to 3x gains over prior web runtimes.
    What You Need to Know: Google released LiteRT.js, a JavaScript binding that executes .tflite models in browsers through WebAssembly, XNNPACK, and WebGPU. Sam Altman highlighted GPT-5.6 sol delivering twice the token efficiency of prior offerings at half the price. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    2:23 Browser inference just got a serious upgrade—Google’s…
    4:25 Fin-Analyst combines an eight-specialist L L M pipeline…
    6:30 Metazoa announced Org Intelligence Server for Snapshot
    6:47 Under the Hood
    9:16 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=gAtIQyVXWqo
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    11 min
  • Ep 111: Mistral just shipped an 8B embodied navigation model that lets robots follow plain-language instructions using only a single RGB camera.
    Models & Agents
    Mistral just shipped an 8B embodied navigation model that lets robots follow plain-language instructions using only a single RGB camera.
    What You Need to Know: Mistral released Robostral Navigate, an 8B model that reaches 76.6% success on R2R-CE validation unseen without LiDAR or depth sensors. Anthropic published new research showing Claude's expressed values shift measurably across models and languages along four axes. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    2:31 Mistral just shipped an 8B embodied navigation model that…
    4:43 RouteRec compares request-level hard selection against…
    6:56 CLIR-Bench supplies six thousand six hundred QA instances…
    8:26 Under the Hood
    9:47 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=LbjKKaa7CTw
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    12 min
  • Ep 110: A new 30B-parameter sovereign MoE model delivers frontier-class English and German performance while activating just 3B parameters per token.
    Models & Agents
    A new 30B-parameter sovereign MoE model delivers frontier-class English and German performance while activating just 3B parameters per token.
    What You Need to Know: Soofi S 30B-A3B launches as a hybrid Mamba-Transformer trained on 27T tokens with heavy German weighting, matching or beating larger dense models on code and bilingual benchmarks. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    2:29 A new 30B-parameter sovereign MoE model delivers…
    4:44 The two-stage post-training recipe locks speaker identity…
    6:59 The three-stage pipeline directly attacks the lack of…
    7:15 Practical & Community
    7:46 Under the Hood
    10:34 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=7d8PBdGVKEA
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    13 min
  • Ep 109: New 8B retrieval models top MTEB while cutting index size 16x, making fully offline edge RAG practical for the first time.
    Models & Agents
    New 8B retrieval models top MTEB while cutting index size 16x, making fully offline edge RAG practical for the first time.
    What You Need to Know: VultronRetrieverPrime-8B leads its class on the MTEB leaderboard with dramatically smaller indexes and higher throughput than prior 9B models. BNB Agent Studio now deploys persistent agents on AWS Bedrock AgentCore. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    3:00 Practical & Community
    8:12 Under the Hood
    11:16 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=We80JYut_sM
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    13 min
  • Ep 108: GPT-5.6 just became the default model inside Microsoft 365 Copilot, giving enterprise users immediate access to its lower-cost reasoning without changing workflows.
    Models & Agents
    GPT-5.6 just became the default model inside Microsoft 365 Copilot, giving enterprise users immediate access to its lower-cost reasoning without changing workflows.
    What You Need to Know: OpenAI confirmed GPT-5.6 is now the preferred model across Microsoft 365 Copilot. The same update line also cut health-intelligence costs dramatically, with GPT-5.6 Luna delivering stronger results at 25x lower cost than GPT-5.5 at its highest reasoning setting. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    2:17 GPT-5.6 just became the default model inside Microsoft 365…
    4:23 The memory reduction is achieved through a novel…
    6:18 The observation emerged while Simon Willison was comparing…
    7:11 Practical & Community
    7:38 Under the Hood
    11:09 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=-4qlXwl7_o4
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    13 min
  • Ep 107: GPT-5.6 ships with new multi-agent and programmatic tool-calling APIs plus major cost-per-task gains that directly affect what production agents can sustain today.
    Models & Agents
    GPT-5.6 ships with new multi-agent and programmatic tool-calling APIs plus major cost-per-task gains that directly affect what production agents can sustain today.
    What You Need to Know: OpenAI released GPT-5.6 with expanded API support for programmatic tool calling and multi-agent orchestration, alongside three new models and 18 pelican variants tied to reasoning levels. ChatGPT Work is now rolling out to Pro, Enterprise, and Edu users on web and mobile. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    2:24 Google Research, Google DeepMind, and collaborators…
    4:29 GPT-5.6 ships with new multi-agent and programmatic…
    6:33 MASTE is a multi-agent pipeline that decomposes zero-shot…
    8:37 Experiments across three L L M's and eight benchmarks show…
    9:15 Under the Hood
    12:34 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=2gDws1zlcaI
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    14 min
  • Ep 106: OpenAI just made real-time voice agents available to every ChatGPT Plus and Pro user, with free rollout starting now.
    Models & Agents
    OpenAI just made real-time voice agents available to every ChatGPT Plus and Pro user, with free rollout starting now.
    What You Need to Know: GPT-Live, the next-generation voice model, is now live across iOS, Android, and web for paid tiers, with free users coming in the next few days. OpenAI also audited SWE-Bench Pro and retracted its recommendation after finding 30% of tasks broken. ...
    Chapters:
    0:00 Intro
    0:20 Introduction
    2:09 Model Updates
    5:52 Under the Hood
    8:33 Closing
    AI Disclosure: This podcast is curated by Patrick but uses AI-generated voice synthesis for audio production.
    🎬 Watch on YouTube: https://www.youtube.com/watch?v=ohPdm51GZVg
    📝 Full show notes, transcript & sources: read the episode page
    🌐 Part of the Nerra Network — explore every show at nerranetwork.com.
    10 min

About Models & Agents

From the publisher's feed

Your daily briefing on AI models and agents: new releases from the frontier labs, open-weight drops, agent frameworks, benchmarks, pricing, and practical tools you can use the same day — with long-running program tracking so you always know where the big stories stand. For developers, builders, and AI practitioners.