Inworld launched Inworld Inference, an LLM inference API built for consumer AI apps, serving open models including Gemma 4, GLM, DeepSeek, and MiMo.

> Inworld optimizes how each model runs to deliver faster responses for tutors, companions, games, and other consumer apps.

> On Gemma 4 26B, median time to first token is 0.17 seconds, compared with 0.61 seconds for the median of the other providers on OpenRouter - 3.6× faster.

> At the 95th percentile, time to first token is 0.31 seconds versus 2.42 seconds for the median of those providers - 7.8× faster.

> Developers can also set requests to switch automatically to a fallback model if the first model fails or takes too long.

Gemma 4 26B costs $0.07 per million input tokens - nearly 40% below the median among the compared providers offering the same model that don't train on customer data.

Developers can choose a service tier for each request, prioritizing live replies while spending less on tasks that can wait.

Try hosted models in the playground!
https://platform.inworld.ai
#sponsored

Claude Managed Agents, the Messages API, and the Claude SDK now support OpenTelemetry traces.

> By default, they include duration, status, token counts, model, and tool names.

Anthropic is steadily enhancing its cloud platform, and supporting OpenTelemetry is a major step toward enabling AI observability with tools like Grafana, Splunk, and others.

Traces are useful for testing too! 👀

NVIDIA 🔥: A new Nemotron 3.5 Super VL model has been announced!

NVIDIA Nemotron 3.5 Super VL is an open model with 120B total parameters and 12B active.

The model supports reasoning, tool use with text, image, document, and video understanding for long-running AI agents.

Krea launched Krea 3 for fast and precise image editing

Krea 3 is a new image-editing model for prompt-based 2K edits in under 8 seconds, with region control, bounding boxes, multi-turn edits, LoRA fine-tuning, 4K support, and API access for design, marketing, and commerce teams.

🗞 #sponsored @testingcatalog

META 🔥: Muse Spark Open (Muse-Spark-Open-816B-A42B) is now available on HuggingFace!

> "Muse Spark Open is a mixture-of-experts model with 42B active parameters, 816B total parameters, and a 1M-token context window."

> "It’s built for agentic and coding workflows, from understanding large repositories to completing multi-step engineering tasks with tools."

Benchmarks 👀

MICROSOFT 🔥: MAI-Base-2 pre-training has been completed; MAI-Base-3 training cycle is already being prepared as well.

> "MAI-Base-2 outperforms every open-weight model we tested on our held-out evaluations. It is 4-12x more pre-training FLOPs-efficient than the strongest open base model, DeepSeek-V4-Pro-Base (see Figure 3 below), and we have put that saving into training a larger model."

* Figure 1 shows the performance of MAI-Base-2 and other models with at least one trillion total parameters across a range of pre-training evaluations (lower is better).

BREAKING 🔥: OpenAI shipped an upgrade to its Voice Mode on ChatGPT, Codex and dots!

A new model, GPT-Live-1.1, comes with support of new voices in French, German, Hindi, Mexican Spanish, Korean, and Japanese.

A constant “Checking..” should be gone as well. It has been quite annoying during vibe coding sessions.

API support is coming soon 👀

Black Forest Labs launched FLUX 3 Image Klein, its new open-weight image model!

> "Customizable and fast. 5.6B parameters and runs on a single consumer GPU. Native 4K, bounding boxes, and pixel-perfect editing."

> Developers can post-train and fine-tune FLUX 3 Image Klein.

> It comes with a low-latency version for faster generation and self-hosting, as well as quantized versions - FP8, NVFP4, and INT8.

> The distilled model with native 4K will also be available on BFL APIs.

DAILY AI BRIEF 🗞 — Oct 16

META 🔥:
- Meta released Muse Spark Open, an open-weight 816B-total, 42B-active MoE model with a 1M-token context window, on Hugging Face and optimized to run locally on Mac Studio.

NVIDIA 🔥:
- NVIDIA released Nemotron 3.5 Super VL, an open 120B-total, 12B-active multimodal model for long-running AI agents.
- NVIDIA also introduced Nemotron 3.5 Bolt 24GB, a model that fits into 24GB of memory.

OPENAI 🔥:
- OpenAI is rolling out GPT-Live-1.1 to Voice in ChatGPT, Codex and dots for all users, with new voices in six languages and API support coming soon.
- OpenAI launched the Rosalind Academic Plan, giving eligible US university labs sponsored access to GPT-Rosalind.
- Sites is now generally available for ChatGPT Enterprise with US and EU data residency.
- ChatGPT can now view your open Chrome tabs.
- Codex landed changes and more compute to sharply reduce "Selected model is at capacity" errors.

GOOGLE 🔥:
- Gemini 4 Argon is rolling out in Antigravity to paid Google AI Pro subscribers with a separate weekly Argon quota.
- Google launched Universal Cart, a cross-retailer shopping cart, in the US.
- Circle to Search can now check whether a text message is a scam.

MICROSOFT 🔥:
- Microsoft finished pre-training MAI-Base-2 and is already preparing the MAI-Base-3 training cycle.
- GitHub made agentic autofix for code scanning alerts generally available.

ANTHROPIC 🔥:
- Anthropic gave a team from METR employee-like access to assess its safety practices.

PERPLEXITY 🔥:
- Perplexity open-sourced SPACE, the Firecracker-based sandbox that runs every Perplexity Computer task, under Apache 2.0.
- Nemotron 3.5 Super VL is now in Perplexity for Pro and Max subscribers and in the Agent API.

CURSOR 🔥:
- Cursor added Muse Wonder at $4/$20 per million input/output tokens, scoring 46.1% on CursorBench.

BLACK FOREST LABS 🔥:
- Black Forest Labs released FLUX 3 Image Klein, a 5.6B open-weight image model with native 4K that runs on a single consumer GPU.

KREA 🔥:
- Krea launched Krea 3, an image-editing model for prompt-based 2K edits in under 8 seconds.

AMAZON 🔥:
- Amazon's Bee launched two new personal AI devices starting at $75.

Imagine testing this for the 1st time 🥵