NVIDIA Nemotron 3.5 Super VL is an open multimodal model that runs with end-to-end encrypted inference at over 500 output tokens per second on a single NVIDIA B300 GPU. It reads documents, screenshots, images and video, reasons over them and calls tools through an OpenAI-compatible API. The model uses a 120B-parameter mixture-of-experts architecture with only 12B active per token.
Key points
500+ output tokens per second for a single request on one B300 GPU.
120B total parameters, 12B active MoE per token, with built-in Multi-Token Prediction.
Sources
GPT-6.1 Sol uses 53% fewer credits than Gemini 4 Argon
GPT-6.1 Sol uses about 53% fewer credits than Gemini 4 Argon in a Three.js 3-D shooter test. Gemini 4 Argon runs faster, is more token efficient, and generated three weapons and three enemy types.
Key points
Sol: ~53% fewer credits
Argon: faster, more token efficient; 3 weapons, 3 enemy types
Sources
Risphai web search cheaper and faster, uses 2–6 credits
Risphai web search is now cheaper and faster, using 2–6 credits per query instead of 3.5–7.5 credits. The change was announced in a tweet. Powered by @ExaAILabs.
Key points
Credits per search: 2–6 (previously 3.5–7.5)
Announcement made via quoted tweet stating the update
Sources
EffectTS LSP and Oxlint rules faster, at least 15% improvement
EffectTS LSP and Oxlint rules are now faster, delivering at least a 15% improvement when run in patched tsgo mode or oxlint with typecheck:true. Additional optimisations are planned for the future.
Key points
Faster LSP and Oxlint rules
At least 15% improvement in patched tsgo mode or oxlint with typecheck:true
Sources
Reflection Beam gets Day 0 support on Ray Serve LLM
Reflection Beam model receives Day 0 support on Ray Serve LLM. Beam is an open-weight, 500B-parameter sparse MoE model optimized for coding and agentic workloads.
Key points
Day 0 support for Reflection's Beam model on Ray Serve LLM.
Beam is a sparse MoE model in the 500B-parameter class, built for coding and agentic workloads.
Sources
Prismor adds mirror, Claude Code blocks, 5 new agents
Prismor open source runtime control plane introduces a mirror that runs agent tools, checks each call and masks secrets in output. It also shows blocks live inside Claude Code right under the prompt and adds governance for five additional agents.
Key points
Mirror runs agent's tools, checks every call and masks secrets in the output
Blocks show up live inside Claude Code, right under the prompt
Governance extended to five more agents
Sources
- Original post
- .prismor/policy.yaml - Linked page
Opus 5.5 wildfire analysis prototype maps spread
Opus 5.5 wildfire analysis prototype maps fire’s spread over time and tracks hot spots. It is a recent project highlighted among six.
Key points
Prototype maps fire spread over time.
Prototype tracks hot spots, built with Opus 5.5.
Sources
IFS Intelligence announcement at IFS Unleashed
IFS Intelligence was announced at IFS Unleashed event. Zach, the engineering lead, discussed building it with Harshana from IFS as a joint effort across two companies.
Key points
Announcement: IFS Intelligence announced at IFS Unleashed.
Collaboration: Zach, eng lead, talked with Harshana from IFS about building it as one team across two companies.
