Pika H3 Flex debuts as the #1 open weight model on DesignArena’s t2v board, offering exportable audio stems and up to 4k resolution. It generates a 15-second clip with soundtrack in as little as 7 seconds and runs faster than real time.
Key points
Rank: 1 open weight model on DesignArena’s t2v board
Speed: 15-second clip with soundtrack produced in 7 seconds, faster-than-realtime generation
Sources
GeneralMagic AI 2x more accurate, 24 s faster
GeneralMagic AI was 2x as accurate and 24 seconds faster on average than 7 frontier models and 2 enterprise assistants in an underwriting benchmark. The test used the same insurance documentation for all systems.
Key points
Benchmark: compared against 7 leading frontier models and 2 enterprise AI assistants using the same insurance documentation
Result: 2x as accurate and 24 seconds faster on average than the other AI systems tested
Sources
Dream introduces new models and raises $17M seed
Dream announced a new class of models that reason about physical space, geometry, constraints, and object-space relationships. The startup disclosed $17 million in seed funding and targets the $13 trillion architecture, engineering, and construction industry.
Key points
Funding: $17M seed from a16z, svangel, xyz_vc, and ziggcap.
Focus: models trained to reason about physical space, its geometry, constraints, and object-space relationships.
Sources
Cohere Series E funding up to CAD $7B at $35.7B valuation
Cohere announced the initial close of its Series E, targeting up to CAD $7 billion and a CAD $35.7 billion valuation. The round is described as the largest private financing round in Canadian startup history.
Key points
Series E initial close targeting up to CAD $7 billion
Valuation set at CAD $35.7 billion
Largest private financing round in Canadian startup history
Sources
Atelico AI Engine launches on-device game AI, 121 FPS
Atelico AI Engine is a library that runs AI inside games and now achieves 121 FPS while generating text. It bundles language models, image generation, voice, speech recognition, lip-sync, memory, search, guardrails, and a behavior engine. The engine integrates with Godot, Unity and Unreal and supports Windows, macOS, Linux, iOS and Android.
Key points
FPS: first test dropped from 123 FPS to 3 FPS; after engine integration the same scene holds 121 FPS while a model generates text.
Compatibility: plugs into Godot, Unity and Unreal, has a C ABI, and runs on Windows, macOS, Linux, iOS and Android.
Sources
- Original post
- atelico.studio - Linked in the post
- atelico.studio - Linked in the post
Le Chonk weights released; ranks #1 on FrontierCode
Le Chonk weights released; ranks #1 on FrontierCode and sets a new Pareto efficiency benchmark on artificialanalysis. Positioned for cyberdefense, tested by CrowdStrike, Cloudflare, Rubrik, Zscaler and others, and available on HuggingFace for customization with Mistral Forge.
Key points
Rank: 1 on @cognition’s FrontierCode among open models
Benchmark: new benchmark for Pareto efficiency on @artificialanalysis
Testing: evaluated by CrowdStrike, Cloudflare, Rubrik, Zscaler and others
Availability: download weights on HuggingFace; customize with Mistral Forge
Sources
Relics – new domain-specific swarm configurations announced
Relics introduces domain-specific configurations that change how an entire agent swarm operates. Installing a Relic gives a swarm the methods, infrastructure, and research machinery of that field. Available now to enterprise ArkenOS customers and rolling out to all users soon.
Key points
Relics package all skills, tools, environments, hooks, evals, verifiers, and guardrails into domain-specific configurations.
Installing a Relic makes a swarm inherit the methods, infrastructure, and research machinery of that field.
Sources
Vercel agents can send feedback via Vercel MCP
Vercel agents can send feedback via Vercel MCP, attaching structured context to comments. Use plain language to describe bugs or feature requests; the system calls submit_feedback.
Key points
Capability: agents can send feedback to Vercel.
Method: submit_feedback via Vercel MCP attaches structured context and submits the comment.