Trump Joins All-In Summit by Speakerphone to Say Robots Won't Take Over — theverge.com

Nvidia's Huang and Trump align on AI pace at All-In Summit — techcrunch.com

Cornelis secures $205M to challenge Nvidia's grip on AI networking infrastructure — techcrunch.com

Valve Steam Frame price settles at $1,059 as component costs squeeze margins — theverge.com

Mushroom Coffee Roundup: Wired Names Atlas Latte Superblend the Top Pick — wired.com

Trump Joins All-In Summit by Speakerphone to Say Robots Won't Take Over — theverge.com

Nvidia's Huang and Trump align on AI pace at All-In Summit — techcrunch.com

Cornelis secures $205M to challenge Nvidia's grip on AI networking infrastructure — techcrunch.com

Valve Steam Frame price settles at $1,059 as component costs squeeze margins — theverge.com

Mushroom Coffee Roundup: Wired Names Atlas Latte Superblend the Top Pick — wired.com

Hardware

RTX Pro 5500 Blackwell lands with 84GB of GDDR7 for one-pool local LLMs

NVIDIA listed the RTX Pro 5500 Blackwell Workstation Edition with 21,760 CUDA cores - the same GB202 core count as the GeForce RTX 5090 - plus 84GB of ECC GDDR7 at roughly 1,400 GB/s on a 448-bit interface. The trade is capacity over bandwidth: 28 clamshell GDDR7 chips double the memory of the 5090, which keeps 1,790 GB/s on 512-bit. MIG support means the 84GB can run as one pool for 70B-plus models or split into two isolated 42GB instances.

hothardware.com · 2026-09-14 · Read the dispatch →

llama.cpp 0.4.1 adds Maple 20B-A1B, Tencent Hy 4 and Spark2.5 support

Stable release with better JSON-schema handling, chat parsing and server child-process management, a ggml bump to v0.24.0, and a new --fuse-qkv flag in nightlies b10969/b10970.

GitHub · 2026-09-14

Ollama 0.34.0 serves local models straight inside ChatGPT Desktop

Adds OpenAI-compatible client tool search, response compaction and faster structured output on Apple Silicon, with no breaking steps for existing Compose or Modelfile setups.

Releasebot · 2026-09-14

Temporal raises $550M Series E at a $12.55B valuation

Lightspeed-led round with Wellington, Goldman Sachs Alternatives and Tiger Global co-leading; durable execution infrastructure for agentic workloads, with OpenAI, Snap, Nvidia and JPMorgan Chase as customers.

Unite.ai · 2026-09-14

Research & Papers

When agents slow down: an Elo-per-token read on test-time strategies

Normalises agent accuracy by tokens spent across revision, tool use and stop decisions, giving a directly usable efficiency measure for latency- and VRAM-constrained local agent setups.

HuggingFace · 2026-09-14

Creative exploration for vibe-design agents gets its own scaffold

Separates exploration from implementation in natural-language-to-UI agents so candidate interfaces diversify without the blunt effect of raising token-level temperature, while syntax-sensitive code stays stable.

HuggingFace · 2026-09-14

Omni-Streaming Thinking fixes premature cross-modal commitment

Identifies a failure mode where an early visual interpretation persists after contradicting audio arrives, and updates commitments as video chunks and synchronized audio are processed.

HuggingFace · 2026-09-14

Learned poison-set selection strengthens LLM backdoors

Random sampling of a fixed number of poisoned examples underestimates worst-case vulnerability; across three LLaMA-3-8B settings, learned selection beats random at equal budget.

HuggingFace · 2026-09-14

Atria Dawn previews an agentic foundation model for research work

A foundation agentic language model trained through a verified pipeline for scientific and engineering workflows, where agents also take part in developing their successors.

HuggingFace · 2026-09-14

Open Weights & Models

Edge0-35B-A3B-preview: 35B MoE in under 3 GiB via SSD expert streaming

A 35B mixture-of-experts checkpoint (~3B active per token) shipped as 4-bit with LoRA and prerouter adapters; weights stay on disk and only the needed experts are streamed, so the full model fits phone-class memory budgets. Apache-2.0 safetensors on the Qwen3.5 MoE line.

HuggingFace · 2026-09-14

colibri: a pure-C inference engine for frontier MoE models

Number one on GitHub trending today: a dependency-free C engine that runs 744B to 2.8T parameter MoE models on consumer and heterogeneous hardware by treating storage, RAM and VRAM as one inference hierarchy and streaming experts from disk.

GitHub · 2026-09-15

Agnes-3.0-Flash lands on the Hub as a 33B multimodal checkpoint

Image-text-to-text open-weight release added to the trending list on 2026-09-14.

HuggingFace · 2026-09-14

VoiceStudio: a fully-local ElevenLabs-style voice stack

Open-source alternative covering voice cloning, voice design, video dubbing, dictation and transcription; +2,776 stars today.

GitHub · 2026-09-15

Tools & Runtimes

Pi-box ships a coding agent as a VM inside WASM inside Electron

Self-contained local coding agent packaged as an Electron app, with the agent running in a WASM-hosted VM; repository is open source.

GitHub · 2026-09-15

AgenticOS is an open-source control plane for company AI agents

Show HN launch of a control plane for coordinating multiple AI agents, with the code published on GitHub.

GitHub · 2026-09-14

VideoHighlighter analyzes video offline and self-hosted

Show HN utility that runs without external API dependencies.

GitHub · 2026-09-14

Self-Hosted Stack

n8n Assistant is live for Cloud and self-hosted Docker installs

The Assistant plans, builds, runs and debugs workflows from plain-language prompts on the canvas, asks for credentials only when needed and gates activation behind user approval; still in preview, with plan AI credits separate from Gateway credits.

Unroll Now · 2026-09-14

n8n idempotency keys: which primitives actually dedupe retries

Keys are effectively undeclared, so operators should reuse one key per retried operation, cap retries and verify a webhook delivery ID first; the practical primitives are execution.id as stable identifier, the Data table node as a dedup log and Error Trigger for failed runs.

apis.io · 2026-09-14

Hardware & Edge

Perplexity Portable Computer reaches Windows on 24GB-plus RTX cards

The local-first agentic Portable Computer now runs on GeForce RTX and RTX PRO hardware after Linux and DGX Spark, with acceleration requiring at least 24GB of VRAM.

pcguide.com · 2026-09-14

Arm opens an AI Portal for on-device agentic AI across CPUs, GPUs and NPUs

The portal publishes pre-optimized Qwen, Gemma and YOLO builds with latency, memory and size data for ExecuTorch, LiteRT and ONNX-RT, and points to Arm CSS for Mobile 2 running models on SME2 and GPU neural accelerators.

design-reuse.com · 2026-09-14

NVIDIA and Palantir ship a sovereign AI stack that runs on-premises

Pairs NVIDIA Nemotron open models with Palantir Foundry, AIP and Ontology for on-prem, colocation or cloud deployment, using cuOpt, NeMo Data Libraries and NeMo AutoModel for optimization and scenario planning.

eenewseurope.com · 2026-09-14

Free Models

Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.

In Brief

Source:

RTX Pro 5500 Blackwell lands with 84GB of GDDR7 for one-pool local LLMs

llama.cpp 0.4.1 adds Maple 20B-A1B, Tencent Hy 4 and Spark2.5 support

Ollama 0.34.0 serves local models straight inside ChatGPT Desktop

Temporal raises $550M Series E at a $12.55B valuation

Research & Papers

When agents slow down: an Elo-per-token read on test-time strategies

Creative exploration for vibe-design agents gets its own scaffold

Omni-Streaming Thinking fixes premature cross-modal commitment

Learned poison-set selection strengthens LLM backdoors

Atria Dawn previews an agentic foundation model for research work

Open Weights & Models

Edge0-35B-A3B-preview: 35B MoE in under 3 GiB via SSD expert streaming

colibri: a pure-C inference engine for frontier MoE models

Agnes-3.0-Flash lands on the Hub as a 33B multimodal checkpoint

VoiceStudio: a fully-local ElevenLabs-style voice stack

Tools & Runtimes

Pi-box ships a coding agent as a VM inside WASM inside Electron

AgenticOS is an open-source control plane for company AI agents

VideoHighlighter analyzes video offline and self-hosted

Self-Hosted Stack

n8n Assistant is live for Cloud and self-hosted Docker installs

n8n idempotency keys: which primitives actually dedupe retries

Hardware & Edge

Perplexity Portable Computer reaches Windows on 24GB-plus RTX cards

Arm opens an AI Portal for on-device agentic AI across CPUs, GPUs and NPUs

NVIDIA and Palantir ship a sovereign AI stack that runs on-premises

Repos on GitHub

Trending on HF

Free Models

Free on OpenRouter

Free on OpenCode Zen

In Brief