Nvidia reintroduces DLSS 5 neural rendering after a March backlash — theverge.com

Nvidia’s DLSS 5 launches this week with heavy GPU demands — theverge.com

Running a Chatbot on Your Own Computer — wired.com

Nvidia CEO Took Trump’s Call Mid-All-Hands, Sources Say — wired.com

Nvidia reintroduces DLSS 5 neural rendering after a March backlash — theverge.com

Nvidia’s DLSS 5 launches this week with heavy GPU demands — theverge.com

Running a Chatbot on Your Own Computer — wired.com

Nvidia CEO Took Trump’s Call Mid-All-Hands, Sources Say — wired.com

Spark chips, deepened

NVIDIA and MediaTek pledge multiple generations of RTX Spark and DGX Spark PC chips

NVIDIA and MediaTek are expanding their partnership to build several future generations of RTX Spark and DGX Spark PC chips, pairing NVIDIA GPUs (including Vera Rubin and upcoming Feynman architectures) with MediaTek SoCs for consumer PCs and workstations. MediaTek also adopts NVIDIA's NVLink Fusion for custom AI infrastructure. A multi-generation commitment to affordable, local AI hardware.

MediaTek · 2026-08-31 · Read the dispatch →

Nvidia's RTX 5090 now costs at least $5,000 as Blackwell price hikes bite

A freshly bought RTX 5090 runs at least $5,000 in North America — roughly 2.5x its $1,999 launch price — with the wider RTX 50-series up 20-77%. Raises the bar for building VRAM-capable local inference rigs.

Tom's Hardware · 2026-09-01

Slotstream runs Qwen3.8-Flash-Next 4-bit on low-RAM Macs by streaming experts from SSD

A new MLX + Swift tool streams MoE expert weights from SSD instead of holding them in RAM, running the 125B Qwen3.8-Flash-Next at 4-bit on Macs with far less memory, and exposing an Ollama-compatible API.

GitHub · 2026-08-31

ASUS RTX Spark laptops sell out channel pre-orders ahead of Q4 launch

ASUS confirmed initial RTX Spark ProArt P16/P14 allocations are completely sold out via channel pre-orders, ordering more chips ahead of the planned launch. The RTX Spark SoC offers up to 128GB unified memory, running models up to ~120B parameters locally.

Yahoo Tech · 2026-08-31

Research & Papers

A universal context-reuse layer for cross-model KV sharing

Proposes a universal layer that reuses KV caches across different LLMs, eliminating redundant prefill when multiple models share context and cutting memory and latency in serving.

arXiv · 2026-08-31

Q-Strata: hierarchical bit allocation for mixed-precision MoE quantization

Assigns per-layer bitwidths across every expert under a fixed budget, minimizing quality loss in mixed-precision quantization of MoE LLMs.

arXiv · 2026-08-31

NN-PPI calibrates small language models for claim check-worthiness at lower cost

A pointwise Prediction-Powered Inference extension lets small SLMs recover LLM-level accuracy for claim check-worthiness at a tenth of the serving cost; residual calibration complements fine-tuning.

arXiv · 2026-08-31

DASC: decay-aware state compression for hybrid linear-attention serving

Handles in-place recurrent-state updates that complicate prefix reuse and KV-cache management in hybrid linear-attention models, cutting cache growth.

arXiv · 2026-08-31

Sliding-window attention beats linear attention for long contexts

Microsoft report showing sliding-window attention matches or beats linear attention for long-context efficiency, informing architecture choices that reduce memory and inference cost on local hardware.

HuggingFace · 2026-08-31

Open Weights & Models

NVIDIA quantizes Qwen3.8-2.4T-A95B to 4-bit NVFP4 — roughly 2.5x smaller on vLLM

NVIDIA Model Optimizer v0.46 NVFP4 quant drops bits-per-parameter from 16 to 4 with negligible accuracy loss, cutting disk and VRAM ~2.5x for Blackwell via vLLM/SGLang.

HuggingFace · 2026-08-27

Unsloth ships 1-3-bit GGUFs of GLM-5.3-Flash — 320B/18B now runs on 128GB RAM

Unsloth quantized Z.ai's GLM-5.3-Flash down to dynamic 3-bit GGUFs that fit a 128GB Mac or workstation, with 63.7k downloads already; needs Unsloth's llama.cpp branch until mainline merges support.

HuggingFace · 2026-08-26

Tools & Runtimes

shaide: K8s-native AI platform for distributed multi-model inference

An open-source Kubernetes-native platform pooling CPU/GPU resources for distributed multi-model inference across a cluster, targeting self-hosted production serving.

GitHub · 2026-09-01

llama.cpp b10729 adds Metal fa-vec tunings for M1 Ultra

The runtime's steady patch stream continues with Apple M1 Ultra Metal backend tunings, plus a follow-up adding qwen4exp indexer-head slicing.

GitHub · 2026-08-31

Open WebUI v0.11.3: chat-branch reload fix, upgrade safety and MCP OAuth fix

Patch release fixing chat branches falling out of view after reloads, stopping half-updated database upgrades after migrations, and restricting MCP tool-server OAuth controls.

GitHub · 2026-08-31

extensible-mcp: a policy proxy between LLMs and MCP servers the model cannot reach

An open-source proxy enforcing policy the model itself cannot read or bypass, plus tools to launder requests — a control point for governance on local MCP toolchains.

GitHub · 2026-09-01

PromptLab: local-first Chrome side panel for LLM prompt testing

A local-first Chrome side panel for testing and iterating on LLM prompts directly in the browser, without cloud round-trips.

GitHub · 2026-09-01

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:

NVIDIA and MediaTek pledge multiple generations of RTX Spark and DGX Spark PC chips

Nvidia's RTX 5090 now costs at least $5,000 as Blackwell price hikes bite

Slotstream runs Qwen3.8-Flash-Next 4-bit on low-RAM Macs by streaming experts from SSD

ASUS RTX Spark laptops sell out channel pre-orders ahead of Q4 launch

Research & Papers

A universal context-reuse layer for cross-model KV sharing

A universal context-reuse layer for cross-model KV sharing

Q-Strata: hierarchical bit allocation for mixed-precision MoE quantization

NN-PPI calibrates small language models for claim check-worthiness at lower cost

DASC: decay-aware state compression for hybrid linear-attention serving

Sliding-window attention beats linear attention for long contexts

Open Weights & Models

NVIDIA quantizes Qwen3.8-2.4T-A95B to 4-bit NVFP4 — roughly 2.5x smaller on vLLM

NVIDIA quantizes Qwen3.8-2.4T-A95B to 4-bit NVFP4 — roughly 2.5x smaller on vLLM

Unsloth ships 1-3-bit GGUFs of GLM-5.3-Flash — 320B/18B now runs on 128GB RAM

Tools & Runtimes

shaide: K8s-native AI platform for distributed multi-model inference

shaide: K8s-native AI platform for distributed multi-model inference

llama.cpp b10729 adds Metal fa-vec tunings for M1 Ultra

Open WebUI v0.11.3: chat-branch reload fix, upgrade safety and MCP OAuth fix

extensible-mcp: a policy proxy between LLMs and MCP servers the model cannot reach

PromptLab: local-first Chrome side panel for LLM prompt testing

Repos on GitHub

Trending on HF

Free Models

Free on OpenRouter

Free on OpenCode Zen

In Brief