TL;DR
6 takeawaysElorian, the startup that raised a $55 million seed round before shipping anything, releases Glimpse, a research-preview model that crops, pulls video frames and runs code while it reasons about images
Sources: TechCrunch
NVIDIA releases Nemotron 3.5 Super VL, an open 120B-parameter multimodal model with 12B active parameters, with open weights, datasets and training recipes for building specialized agents
Sources: NVIDIA
OpenAI releases prompt-to-voice for all API customers and nine new GPT-Live voices in German, French, Hindi, Japanese and Korean
Sources: OpenAI
Short: Vercel data shows the decision model Jev reached 35% of paid teams on its AI gateway within two weeks
Sources: TechCrunch
Short: Reflection releases the Beam open weights, a 501B mixture-of-experts model with 23B active parameters and a 1M-token context, under Apache 2.0
Sources: Hugging Face
Short: Perplexity open-sources SPACE, the Firecracker microVM sandbox platform it says runs its Computer sessions
Sources: PerplexityGitHub
The Details
Story 01 / 03
Elorian debuts Glimpse, a model that looks back at the image while it reasons
- What happened
- Elorian, the Palo Alto startup that raised a $55 million seed round at a $300 million valuation before launching a product, released its first model, Glimpse, on Friday, TechCrunch reports. It is a research preview with API access, and the chat interface is waitlist-only. Glimpse can crop into a region, pull specific frames from a video or run code to measure and compare things in the middle of its reasoning, a mechanism Elorian calls a “visual scratch space.” Pricing is $0.60 per million input tokens and $3 per million output tokens.
- Why it matters
- Most multimodal models look at an image once and then reason in text. TechCrunch notes that OpenAI added crop-and-zoom during reasoning to its o-series models in 2025, but Elorian says it built the whole model around the idea. According to Elorian’s own evaluations, Glimpse outperforms GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results elsewhere. These are company-reported results.
- Enterprise impact
- Teams sitting on large video or chart archives, such as robotics, labs and media, can test the API against their own footage before trusting vendor benchmarks.
- Startup impact
- A small team trained Glimpse on a few hundred GPUs, according to Elorian, which suggests a narrow, well-aimed model can compete in a crowded field.
- What to focus on
- Independent evaluations of Glimpse, and when the waitlisted chat product opens to everyone.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 02 / 03
NVIDIA releases Nemotron 3.5 Super VL for multimodal agents
- What happened
- NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter mixture-of-experts model that activates 12 billion parameters per token. It combines reasoning, instruction following and tool use with understanding of text, images, documents and video. NVIDIA says it is available now on Hugging Face, OpenRouter and ModelScope and as an NVIDIA NIM, and ships with open weights, open datasets and training recipes. NVIDIA names Atlassian, CrowdStrike, Salesforce and Perplexity among the companies integrating or customizing it.
- Why it matters
- NVIDIA says the model matches its largest reasoning model on multimodal agent tasks at roughly one-fifth the parameter count, and is up to five times faster than the next-best open model in its class on the Artificial Analysis leaderboard. Those speed and accuracy figures are NVIDIA’s own claims.
- Enterprise impact
- Because the weights and datasets are open, companies can post-train it on their own policies and run it where their data-residency rules require.
- Startup impact
- Open recipes and RL environments in NeMo Gym lower the cost of building a specialized agent without paying for a closed frontier model.
- What to focus on
- Third-party benchmark runs, and how the document and web-page reading holds up on real enterprise files.
Discuss this story
Comment with your GitHub account. One thread per story.
Story 03 / 03
OpenAI adds prompt-to-voice and nine GPT-Live voices
- What happened
- OpenAI’s release notes say prompt-to-voice is now available to all API customers: developers describe a voice in text, through the API or the voice Playground, and reuse it in GPT-Live sessions. Creating a prompt-generated voice is free, and speech generation uses regular GPT-Live session billing. OpenAI also added nine voices for GPT-Live: three in German, two in French, one in Hindi, two in Japanese and one in Korean.
- Why it matters
- A day earlier, OpenAI says it began rolling out GPT-Live-1.1 and a mini version, GPT-Live-1.1 Quip, in ChatGPT Voice with more natural accents, and added native Decisions integration so an app can act while a user is still speaking. Together, the changes give voice apps more languages and a way to define a brand voice without recordings.
- Enterprise impact
- Support and contact-center teams can prototype a branded voice and regional voices without a separate voice-cloning vendor.
- Startup impact
- Voice-app builders get a free way to create custom voices, with costs tied to normal session billing.
- What to focus on
- Voice quality in non-English languages, and how Decisions behaves in live calls.
Discuss this story
Comment with your GitHub account. One thread per story.
Worth a Glance
Jev reaches 35% of paid teams on Vercel’s gateway
TechCrunch reports Vercel data showing the decision model Jev, published in limited release on September 15, was used by 35% of paid teams within two weeks, nearly double the reach of Sonnet 5.5. Vercel’s report also shows open models now account for more than 70% of tokens through its gateway.
TechCrunch(opens in a new tab)Reflection releases Beam open weights
Hugging Face says Beam is a 501B mixture-of-experts model with 23B active parameters per token, a 1M-token context window and an Apache 2.0 license, focused on coding and agentic use, with day-0 support in Transformers, vLLM and SGLang.
Hugging Face(opens in a new tab)Perplexity open-sources SPACE
Perplexity says SPACE runs agent and user code in isolated Firecracker microVMs behind an HTTP API, and has powered its Computer sessions since July. The code is on GitHub.
Perplexity on GitHub(opens in a new tab)
One Thing Not to Miss Today
Models are learning to look twice and speak in your voice
Vision models that zoom mid-thought and voice models you can describe in a sentence are both moving from demos to APIs. Test them on your own data before relying on a vendor’s benchmark table.