Silicon · 2026-10-20
AMD ships GAIA 0.25, retiring ChatAgent in its local agent framework for Ryzen AI
The Windows Python environment drops from about 0.9 GB to 0.1 GB, and 16 GB PCs no longer run out of memory on the desktop app's first launch.
News
What landed this week
Agents hunt, the editor checks. Every item links its source and says why it matters for people who run this stack.
Silicon · 2026-10-20
Asus ProArt P16 with Nvidia RTX Spark N1X costs $7,000 with 128 GB, per Notebookcheck
Notebookcheck finds the $7,000, 128 GB RTX Spark ProArt P16 weighs under 1.8 kg and matches last year's RTX 5070 model on GPU performance.
Agents · 2026-10-20
Cloudflare releases Clairic, a no-reasoning extraction model, on Workers AI
Billed on input only, Clairic cut a claims pipeline to $6,032 per million claims from $11,740 and to 19.3 s from 61.7 s, per Cloudflare.
Studio · 2026-10-20
fal and Black Forest Labs launch FLUX 3 Max, a 720p video model at $0.0425 per second
fal's post-trained FLUX 3 Max ships text-to-video and image-to-video at 720p, priced at $0.0425 per second until November 7, down from a $0.17 list price.
Stack · 2026-10-20
GitHub retires six Copilot models, including GPT-5.5, GPT-5.4 and Grok 4.5
As of October 19, 2026, six models leave most Copilot experiences, with GPT-6.1 Sol, GPT-6 Luna, Gemini 3.8 Flash and Grok 4.7 named as replacements.
Agents · 2026-10-20
Inception ships Mercury 3, a diffusion LLM with 1M context at $0.20/$0.70 per M tokens
Context grows from 260K to 1M tokens, Inception reports 1,068 tokens/sec on NVIDIA GPUs, and a launch discount halves the price to $0.10/$0.35.
Silicon · 2026-10-19
Five devices over Ethernet and USB-C run one 512-expert model
A user reports 37 to 39 tok/s on Qwen3.8 Flash Next at 60k context across two PCs, one MacBook and two iPhones.
Agents · 2026-10-19
DeepSeek charges $0.75 per 1,000 agent search and fetch calls
The Search and Fetch APIs bill per successful request at 5 QPS per account, with paid metering starting November 1, 2026.
Silicon · 2026-10-19
Modders run the official DLSS 5 model on unsupported games
Modders report up to 90 percent gain over the leaked build, while Nvidia states the new model uses 20 percent less VRAM.
Agents · 2026-10-19
Harness2 adds 11.5 points to agent harnesses without ground truth
Google open sourced a framework that edits an agent's prompt, tools, memory and subagents while the agent works, with gains up to 11.5 points.
Agents · 2026-10-19
Latent reasoning agents cut tokens 35.5% and run 2.6x faster
Jiaru Zou et al. report 10.4% better scores on 11 benchmarks and token cuts of 35.5% to 54.7%, with inference 1.8x to 2.6x faster.
Silicon · 2026-10-19
NVIDIA fits a 74B MoE in 24 GB with native 2-bit weights
Nemotron 3.5 Bolt 74B-A9B runs locally on 24 GB via quantization-aware 2-bit training, with llama.cpp support; the model card lists a 256K context and NVIDIA's OpenMDW-1.1 license.
Studio · 2026-10-19
Fish Audio Drama 3 runs on a 16 GB Radeon
Red Curtain ports the NVIDIA-only 35B MoE to ROCm with experts cached from RAM: 11 GB VRAM, speech slower than realtime.
Silicon · 2026-10-19
SonicBoom beats llama.cpp on decode with no Python in the stack
HardenedLinux's C++ and CUDA runtime posts a 12 to 18% decode lead over llama.cpp on one RTX 3050, with prefill at parity on short prompts and an MLIR planner doing the work.
Silicon · 2026-10-19
A 552B MoE runs on one RTX 4090 at 27 tokens/s
A community engine spreads DeepSeek-V4.1-Flash across VRAM, pinned RAM and NVMe, holding 27 tok/s decode on one 24 GB card.