Beam is now released with open weights under Apache 2.0, powered by DFlash 2 which makes model decode up to 3× faster with zero quality loss. The release includes weights, tokenizer, reference implementation, and BF16, FP8, NVFP4 weight formats plus a DFlash2 drafter.
Key points
Decode speed: up to 3× faster with zero quality loss using DFlash 2.
Release assets: weights, tokenizer, reference implementation, BF16, FP8, NVFP4 weights, and a DFlash2 drafter.
Sources
Dominion Energy adds 100 docked drones, saves hours
Dominion Energy added 100 docked drones to its fleet, letting grid crews see conditions and save about 8-12 hours per event. The drones are used for remote inspection and maintenance of critical infrastructure.
Key points
Fleet expansion: 100 docked drones added.
Time savings: ~8-12 hours saved per event.
Sources
Glimpse visual reasoning model launch with $55M funding
Glimpse visual thinking model is released, backed by $55M funding, and claims competitive frontier performance versus Astra on public visual reasoning benchmarks. It watches video and images and actually reasons about what's happening.
Key points
$55M raised by Google and Apple’s best vision researchers.
Glimpse watches video and images and actually reasons about what's happening.
Sources
- Original post
- help.x.com - Linked in the post
Glimpse visual-thinking model launched, finishes first in puzzle race
Glimpse, Elorian’s new visual-thinking model, launched and finished first in a puzzle race with a time of 73 seconds. It is built to understand images and video and does not generate content.
Key points
Glimpse launched by Elorian, built to understand images and video, does not generate anything.
Puzzle race results: Glimpse 73s, Claude Opus 5.5 83s, Gemini 4 Argon 107s, GPT-6 Astra 138s.
Sources
Mirror open source coding agent launch by Reflection
Mirror is Reflection’s open source coding agent optimized for Beam, announced as available now. Install it today.
Key points
Mirror: open source coding agent optimized for Beam
Install it today with the provided command
Sources
Dino Forcing Flow Models paper announced, claims faster convergence
Dino Forcing Flow Models: Do not denoise what you can predict introduces a new paper that claims denoising images and predicting deterministic information like dino yields faster convergence. The paper is available on arXiv as version 1.
Key points
Paper title: Dino Forcing Flow Models: Do not denoise what you can predict.
Claim: Denoising image and predicting (not denoising) related deterministic information provides a model with faster convergence.
Sources
- Original post
- arxiv.org - Linked in the post
Glimpse visual thinking model launch, frontier
Glimpse is Elorian's first visual thinking model with competitive frontier performance compared to Astra on key public visual reasoning. It detected footwork in a bachata video clip.
Key points
Fact: first visual thinking model
Fact: competitive frontier performance compared to Astra on key public visual reasoning
Fact: caught the footwork in a bachata clip
Sources
Glimpse visual model launches, competitive vs Astra
Glimpse is a visual thinking model that launches with competitive frontier performance compared to Astra on key public visual reasoning benchmarks. It aims to improve handling of moving or angled images.
Key points
First visual thinking model: Glimpse.
Competitive frontier performance versus Astra on key public visual reasoning benchmarks.