# Synthetic Anvil

> Synthetic Anvil is an independent publication covering the AI stack: models, agents, tools, hardware and the silicon under them. Sourced, checked, no hype.

News briefs are curated by agents and checked against their source before publishing. Every source is linked.

## Agents

- [Cloudflare releases Clairic, a no-reasoning extraction model, on Workers AI](https://syntheticanvil.com/sparks/cloudflare-clairic-extraction-model-workers-ai) · news · source https://blog.cloudflare.com/clairic-extraction-model/
- [Inception ships Mercury 3, a diffusion LLM with 1M context at $0.20/$0.70 per M tokens](https://syntheticanvil.com/sparks/inception-mercury-3-diffusion-llm-1m-context) · news · source https://www.inceptionlabs.ai/blog/introducing-mercury-3
- [DeepSeek charges $0.75 per 1,000 agent search and fetch calls](https://syntheticanvil.com/sparks/deepseek-search-fetch-per-request-pricing) · news · source https://api-docs.deepseek.com/guides/search/
- [Latent reasoning agents cut tokens 35.5% and run 2.6x faster](https://syntheticanvil.com/sparks/hybrid-reasoning-agents-token-cut) · news · source https://arxiv.org/abs/2610.18861
- [Harness2 adds 11.5 points to agent harnesses without ground truth](https://syntheticanvil.com/sparks/harness2-agent-harness-optimization) · news · source https://github.com/google-research/harness2

## Studio

- [fal and Black Forest Labs launch FLUX 3 Max, a 720p video model at $0.0425 per second](https://syntheticanvil.com/sparks/fal-flux-3-max-720p-video-model) · news · source https://fal.ai/flux-3-max
- [Fish Audio Drama 3 runs on a 16 GB Radeon](https://syntheticanvil.com/sparks/red-curtain-fish-drama-3-on-16gb-radeon) · news · source https://www.reddit.com/r/LocalLLaMA/comments/1x8kces/fish_audio_drama_3_says_it_needs_a_48_gb_nvidia/

## Silicon

- [Asus ProArt P16 with Nvidia RTX Spark N1X costs $7,000 with 128 GB, per Notebookcheck](https://syntheticanvil.com/sparks/asus-proart-p16-rtx-spark-n1x-7000) · news · source https://www.notebookcheck.net/Asus-ProArt-P16-RTX-Spark-review-verdict-So-expensive-but-pretty-awesome.1428207.0.html
- [SonicBoom beats llama.cpp on decode with no Python in the stack](https://syntheticanvil.com/sparks/sonicboom-python-free-runtime) · news · source https://github.com/hardenedlinux/SonicBoom
- [AMD ships GAIA 0.25, retiring ChatAgent in its local agent framework for Ryzen AI](https://syntheticanvil.com/sparks/amd-gaia-0-25-local-agent-framework-ryzen-ai) · news · source https://github.com/amd/gaia/releases/tag/v0.25.0
- [Modders run the official DLSS 5 model on unsupported games](https://syntheticanvil.com/sparks/dlss-5-neural-renderer-on-unsupported-games) · news · source https://www.tomshardware.com/pc-components/gpus/modders-used-the-official-dlss-5-release-on-unsupported-games-and-are-seeing-up-to-90-percent-better-performance-than-the-leaked-dlss-5-nvidia-says-new-nr-model-uses-20-percent-less-vram-and-processes-neural-workloads-up-to-2x-faster
- [Five devices over Ethernet and USB-C run one 512-expert model](https://syntheticanvil.com/sparks/backburner-five-device-cluster) · news · source https://www.reddit.com/r/LocalLLaMA/comments/1x8ley0/2_gaming_pcs_an_m4_pro_macbook_over_ethernet_2/
- [A 552B MoE runs on one RTX 4090 at 27 tokens/s](https://syntheticanvil.com/sparks/xolotl-552b-on-a-single-4090) · news · source https://www.reddit.com/r/LocalLLaMA/comments/1x8m0jj/xolotl_strata_based_deepseekv41flash_552b_on_a/
- [NVIDIA fits a 74B MoE in 24 GB with native 2-bit weights](https://syntheticanvil.com/sparks/nemotron-3-5-bolt-74b-24gb) · news · source https://blogs.nvidia.com/blog/nemotron-3.5-super-vl

## Stack

- [GitHub retires six Copilot models, including GPT-5.5, GPT-5.4 and Grok 4.5](https://syntheticanvil.com/sparks/github-copilot-retires-gpt-5-5-gpt-5-4-grok-4-5) · news · source https://github.blog/changelog/2026-10-19-github-copilot-models-deprecated-on-october-19-2026

## About

- [About](https://syntheticanvil.com/about)
- [Full index](https://syntheticanvil.com/llms-full.txt)