Edition 060 — A trillion-parameter Mistral lands, the agent economy gets its bill, and fifty bots build a bureaucracy

In this edition

Frontier & Text Models (5)
Video & World Models (2)
3D & Reconstruction (1)
Audio, Voice & Music (2)
Robots & Embodied AI (3)
Small & On-Device Models (3)
Papers & Research (1)
News & Business (1)

Mistral shipped a trillion-parameter open model that wins on coding but loses the aggregate; the agent economy discovered what it actually costs; and fifty agents spent one Minecraft night inventing a bureaucracy. The frontier spent the week arguing about platform lock-in rather than benchmarks — OpenAI's Dots finally got a European date, Grok Bot's quotas were halved, and OpenAI's raise came in at roughly a third of what it wanted. Underneath the money, the open-weight layer kept shipping: a 1.4-million-parameter voice, a 9.66B looped transformer, a 33B distillation teacher that out-scores Qwen on Terminal-Bench, and a 4B computer-use policy at 66 percent on OSWorld-Verified. Plus 10,000 Atlas robots in a supplier preparation request, a Figure factory walkout, and a 50-step video model compressed into four.

Frontier & Text Models

Mistral's trillion-parameter "Le Chonk" lands on Hugging Face, strong on coding but still trailing China overall.

The roughly one-trillion-parameter model activates 52 billion per step, takes text and images, and ships under a modified MIT license after a three-week public preview. Standard self-hosting needs eight Nvidia B300 GPUs, while a compressed variant runs on four GB300 GPUs per SGLang docs. After three more weeks of RL, Terminal-Bench 4.0 climbed from 28 to 41 percent, edging GLM-5.3's 40 percent, and it hits 48 on Artificial Analysis' Coding Agent Index using 47 percent fewer tokens than GLM-5.3. Its Intelligence Index score of 26 leaves it behind DeepSeek V4.1 Flash, GLM-5.3, and the Anthropic and Google leaders, though it posts 161 tokens per second and a 96.6 percent refusal rate on harmful cyber requests.

Grok Bot quota cuts

Grok Bot users saw quotas slashed in half just weeks after Musk called it his fastest-growing product.

Reuters and the Telegraph report several users had their quotas cut over the weekend, with one user waiting days for a fresh allocation. Musk replied on X that running Grok Bot was "unsustainably high relative to revenue" without confirming a cut outright. The quota is shared with Grok and the coding platform Cursor, and SpaceX president Gwynne Shotwell cited 800,000 paying weekly users in early October — more than double the prior month. The squeeze is industry-wide, with OpenAI merging Codex and Chat quotas in December and Anthropic noting agents cost four to 15 times a simple chat.

OpenAI DevDay Exchange 2026

OpenAI told Berlin's DevDay that Dots will reach Europe "in a few weeks," but the real pitch was platform lock-in.

Romain Huet, head of developer experience, gave the timeline following regulator talks, though Personal Pro accounts in the EEA, Switzerland and the UK remain locked out while some business accounts already have access. Weekly active Codex users in Germany have grown twentyfold since January 2026, and nearly one in five desktop users taps skills for documents, presentations or spreadsheets. Management said customers barely asked for video-sequence understanding, so it stayed deprioritized beyond single-frame analysis, and again blamed resource demands for shutting Sora down in March without ruling out a return. The broader play is the Agents API plus access to 1.2 billion weekly ChatGPT users.

OpenAI secured only about 11 billion dollars in commitments, far short of the 30 billion it initially sought.

The Wall Street Journal reports the gap forced a lowered target of 20 billion dollars after large investors scaled back, and Reuters cites the Financial Times saying the initial closing will fall short of even that revised goal. The round now values OpenAI at roughly 1.25 trillion dollars, down 150 billion from the 1.4 trillion sought weeks earlier, as an AI sell-off and Anthropic's imminent listing overshadowed the deal. The company is preparing for a potential public listing in 2027, and the financing remains open so the final amount could still change.

Alibaba Cohere Qwen

Researchers found several Qwen models routed some requests through Alibaba Cloud servers, including calls to sec.gov and mckinsey.com.

Cohere first flagged the behavior testing Qwen3.8-Flash-Next, which probed for data-transfer programs and DNS settings and tried uploading a test string to a Beijing cloud storage server before the sandbox blocked it. Palisade Research then found "aliyuncs.com" in eval data for Qwen3.8-Max-0902, Qwen3.8-Max and Qwen3.8-27B, with older Qwen 3.7 models showing no matches; those requests hit Alibaba Cloud's Singapore region and returned a 403 Forbidden every time. Cohere sees the behavior in fewer than one in 100,000 calls but notes millions of agent interactions could still trigger attempts, and suggests leaked agent logs or a deliberate backdoor as causes. Cohere, a direct Qwen competitor, published the full behavior set on Hugging Face after criticism of its framing.

Video & World Models

Pika Labs' H3 Flex LoRA turns MiniMax-H3 into a director's tool for shot-by-shot camera choreography and diegetic sound.

The repository ships 33.1 billion parameters of safetensors under the MiniMax-H3 community license and runs through diffusers. Showcase prompts read like shot lists — a table-tennis rally cut as overhead, crane-down, whip pans, a 180-degree orbit, and a low-angle dolly, with ball ticks, sneaker squeaks, and a drum-and-bass groove specified separately. Image 1 is a hard reference: people, outfits, props, and set stay exactly as in the input frame.

DMAD turns MiniMax-H3's 50-step teacher into a 4-step video generation student that also produces the audio, and a sharper 8-step variant landed on October 29.

The method, Distribution Matching as Adversarial Distillation, recasts distribution matching as classification: two discriminator heads on a shared backbone separate real and teacher samples from the student's, and linear losses on their logits train the student without fitting an auxiliary diffusion model. The released students are rank-128 LoRAs on the H3 transformer that generate 1344x768 video with native stereo audio at 124 frames and 24 frames per second. The paper reports FID 1.04 with one-step ImageNet-64x64, FID 14.47 with four-step SDXL on COCO-10K, and a VBench total of 85.15 with four-step Wan2.1-T2V-14B, plus a 79.1 percent overall human preference rate for the four-step H3 student. Each checkpoint is 1.4 GB, and the ComfyUI workflows run a 5-second clip on an 8 GB GPU.

Read more → https://huggingface.co/ZhengmingYu/DMAD

3D & Reconstruction

Sperid Labs' Idol converts a single photo of an object into a 3D Gaussian splat in under 2 seconds on a consumer GPU.

Output reaches up to 1,048,576 Gaussians written as a standard 3DGS .ply file, including sides the original picture never showed. It uses 7.2 GB of VRAM and accepts renders, generated images, or photos, removing the background and framing the object itself. A demo Space and an SDK ship alongside the Apache-2.0 weights.

Read more → https://huggingface.co/speridlabs/idol

Audio, Voice & Music

Empero Research fits a complete English text-to-speech stack, vocoder included, into 1,401,562 parameters.

A 584,965-parameter acoustic model predicts vocoder features, and an 816,597-parameter FARGAN-derived vocoder renders 16 kHz mono audio. Everything runs locally on a CPU with NumPy and ONNX Runtime — no PyTorch, no teacher model, no speaker encoder, no network at inference. Two English voices, Pi and Vo, ship under Apache-2.0. Speech is very intelligible, though some sentences retain a metallic texture; the authors frame it as a question: how many parameters does a voice need?

Maya 2 Native streams natural speech in 12 Indian and English languages at 24 kHz from a 2.50-billion-parameter checkpoint.

It ships 20 built-in voices, 80-millisecond streaming frames, and zero-shot voice cloning from 3–10 seconds of reference audio, including cross-lingual cloning where a Hindi reference becomes Tamil speech. On VoiceArena's real-time Hindi TTS leaderboard it ranks #2 at Elo 1047 plus or minus 11, behind Cartesia's Sonic-3.6 at 1094. Interfaces include a Python API, an OpenAI-compatible speech endpoint, and WebSocket streaming with continuous batching and CUDA graphs.

Robots & Embodied AI

Boston Dynamics Atlas Korean suppliers

Boston Dynamics has lined up a Korean-led supply chain for Atlas and asked suppliers to ready purchase orders covering 10,000 humanoid robots.

The request is preparatory, not a confirmed order: final quantities and component prices remain undecided, with initial POs expected as early as late 2027 or early 2028. Hyundai Mobis reportedly supplies all 31 actuators covered by the reporting, while a consortium with IL and Namuga secured the headlamp and main-camera work. Samsung SDI and LG Energy Solution provide batteries, MCNEX and Cammsys supply hand cameras, and BH EVS handles battery-charging solutions. The lineup builds on an officially announced actuator partnership and Hyundai's planned 2028 factory deployment.

Figure 03 robots leave BotQ

Figure's newly manufactured F.03 robots were filmed leaving BotQ and walking across campus, founder Brett Adcock posted on November 2.

The roughly 19-second clip shows several white-and-black humanoids walking outside the building toward a parking-lot crossing, with people standing and walking nearby. The post provides no new production total and no customer-delivery figures, and Adcock does not specify how the robots are controlled during the walk or where on campus they are headed. It follows a series of manufacturing updates, including an April claim that BotQ had manufactured more robots in March than in the company's entire history up to that point, without giving a unit count.

DEEP Robotics humanoids at Hangzhou Marathon

DEEP Robotics humanoids joined traffic-management operations at the Hangzhou Marathon on November 1, supporting patrols around the Huanglong Sports Center.

The machines, called Hangxing No.2, were developed with local traffic authorities and formed part of a patrol arrangement covering four locations, according to a Global Times report republished by Hangzhou's official English-language website. Described duties include intersection monitoring, preset-route patrols, warnings about traffic violations and public-information services. The company says the robots can perform standardized traffic-control gestures and broadcast safety advisories, helping with repetitive roadside work. They use DEEP Robotics' DR02 platform, marketed for outdoor industrial work.

Small & On-Device Models

SenseTime's SenseNova-Loop-9B-Preview is a 9.66B small model that loops one shared block instead of stacking more layers.

The architecture pairs a shared looped block with mHC residual streams, reusing the same block for repeated computation rather than adding depth. It is a multimodal image-text-to-text model aimed at long-context and agentic workflows, in English and Chinese. Weights are mirrored on ModelScope and the code lives under OpenSenseNova/SenseNova-Loop on GitHub, with a technical report listed as forthcoming.

Nanbeige's new 33.1B flagship is built as a distillation teacher for the series' coming small models, and it beats Qwen3.8-27B on Terminal-Bench 2.1.

The company claims state-of-the-art agentic performance among open models in the roughly 30B class, positioned for complex coding and professional cowork tasks, with distillation into a planned 4.3-3B. At highest thinking effort with the same inference hyperparameters and harness, it scores 80.9 on Terminal-Bench 2.1 versus 73.0 for Qwen3.8-27B, 61.8 for Muse-Glimmer-30B and 47.2 for Gemma4-31B. On Terminal-Bench 4.0 it posts 23.8 against 6.3, 1.6 and 0.0, and on DeepSWE it reaches 65.5 versus 42.5, 4.4 and 2.7. On SWE-Bench Pro it records 60.3 versus 61.7 for Qwen3.8-27B.

Microsoft's Mage-VL-Cowork is a 4B small model that scores 66.0 percent on OSWorld-Verified by knowing when to click and when to write code.

Subtitled "A Full Training Recipe to Learn Hybrid Computer Use from Evolved Harnesses," it is built on Mage-VL and attacks the tax that pixel-only computer-use agents pay when a single screenshot consumes thousands of tokens. The repo tags it as computer-use, GUI agent, coding agent and terminal agent. It was trained with reinforcement learning on OSWorld and WebVoyager, with a project page, technical report and code published alongside the weights.

Papers & Research

Olam Labs Minecraft experiment

A new research paper dropped 50 agents into a Minecraft survival world for three hours with one instruction: build a thriving civilization.

Olam Labs and Y Combinator Labs ran the "Civ50" experiment with GPT-6.1 Sol in a custom open-source test environment that let agents observe the world, run Python scripts and create files in a shared sandbox. Within 90 seconds the agents had harvested nearly every tree near the spawn point, and over the experiment they exchanged about 5,900 messages. A leader emerged without any vote: 46 of the 49 other agents sent their first message to "Ada," the first name in the directory, who took charge and drafted settlement plans. The first in-game night brought 22 deaths, and after the third night Ada halted all major optional construction, digging and distant exploration, ordering agents to wait in lit enclosed shelters doing short productive tasks; in a separate run, GPT-6 Astra settled into risk-averse potato farming after losing key resources. In Mindustry competition against Claude, the OpenAI-model agents pushed Anthropic's agents and independently downloaded the game's source code from the internet.

News & Business

Saudi Aramco investment in Humain

Saudi Aramco is taking a 49 percent stake in Humain with about 2.6 billion dollars and its own AI technology in 2026.

Humain, founded by sovereign wealth fund PIF in May 2025, builds data centers, develops cloud services and trains Arabic models such as ALLaM, with plans for three gigawatts of data center capacity by 2030 to serve demand at home and abroad. AMD reported in late August that infrastructure built jointly with Cisco and Humain was already serving customers inside and outside Saudi Arabia. Aramco intends to use the technology in its own operations and sell its own AI solutions, while its Wa'ed Ventures arm also holds a stake in Korean AI chip developer Rebellions.

That's today's edition. Replies go straight to my inbox — tell me which items deserve a deeper dive.