<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
	<channel>
		<title>Sergey Yurkevich's blog</title>
		<link>https://yurkevi.ch/blog/</link>
		<atom:link href="https://yurkevi.ch/blog/feed.xml" rel="self" type="application/rss+xml" />
		<description>Short notes on AI models, tools and research worth knowing about.</description>
		<item>
			<title>Liquid AI releases LFM3, open-weight 2B models including an omni model and decision models</title>
			<link>https://yurkevi.ch/blog/#2026-11-02-liquid-ai-releases-lfm3-open-weight-2b-models-including-an</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-11-02-liquid-ai-releases-lfm3-open-weight-2b-models-including-an</guid>
			<pubDate>Mon, 02 Nov 2026 19:19:36 +0100</pubDate>
			<description>Liquid AI's LFM3 family has six open-weight models of about 2B parameters, including LFM3-Omni-2B and d1-Omni-2B, which returns calibrated decisions without generating text.</description>
		</item>
		<item>
			<title>Speridlabs releases Idol, an open model that turns one image into a 3D Gaussian splat</title>
			<link>https://yurkevi.ch/blog/#2026-11-01-speridlabs-releases-idol-an-open-model-that-turns-one-image</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-11-01-speridlabs-releases-idol-an-open-model-that-turns-one-image</guid>
			<pubDate>Sun, 01 Nov 2026 15:04:03 +0100</pubDate>
			<description>Idol generates 3D Gaussian splats with up to a million Gaussians from a single image, runs in 7.2 GB of VRAM, and ships under Apache-2.0.</description>
		</item>
		<item>
			<title>Underdog's Bark 3.0 hits 4.94% word error in a 163 MB speech model</title>
			<link>https://yurkevi.ch/blog/#2026-10-25-underdog-s-bark-3-0-hits-4-94-word-error-in-a-163-mb-speech</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-10-25-underdog-s-bark-3-0-hits-4-94-word-error-in-a-163-mb-speech</guid>
			<pubDate>Sun, 25 Oct 2026 10:40:38 +0100</pubDate>
			<description>Underdog released Bark 3.0 on October 21: a 163 MB on-device English speech-to-text model that ties NVIDIA's Parakeet at 4.94% average word error rate.</description>
		</item>
		<item>
			<title>Tencent releases WeVisOne-VL, 2B and 4B open-weight perception models</title>
			<link>https://yurkevi.ch/blog/#2026-10-25-tencent-releases-wevisone-vl-2b-and-4b-open-weight</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-10-25-tencent-releases-wevisone-vl-2b-and-4b-open-weight</guid>
			<pubDate>Sun, 25 Oct 2026 10:36:43 +0100</pubDate>
			<description>Tencent's WeChat Vision team published WeVisOne-VL on October 22: 2B and 4B models doing detection, OCR, segmentation and matting, under Apache 2.0.</description>
		</item>
		<item>
			<title>OpenAI publishes 722 math manuscripts</title>
			<link>https://yurkevi.ch/blog/#2026-10-07-openai-publishes-722-math-manuscripts</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-10-07-openai-publishes-722-math-manuscripts</guid>
			<pubDate>Wed, 07 Oct 2026 07:38:25 +0200</pubDate>
			<description>OpenAI released 722 AI-generated math manuscripts on October 6, including matrix multiplication at exponent 9/4 and integer multiplication below n log n.</description>
		</item>
		<item>
			<title>An AI-generated song with an animation Claude Opus 5.5 wrote as code</title>
			<link>https://yurkevi.ch/blog/#2026-09-27-an-ai-generated-song-with-an-animation-claude-opus-5-5</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-09-27-an-ai-generated-song-with-an-animation-claude-opus-5-5</guid>
			<pubDate>Sun, 27 Sep 2026 20:12:29 +0200</pubDate>
			<description>The "I'm Upping My P(doom)" music video pairs an AI-generated track with an animation that Claude Opus 5.5 produced as rendering code rather than pixels.</description>
		</item>
		<item>
			<title>VAST AI releases Mira-Scene, single-image reconstruction of editable 3D scenes</title>
			<link>https://yurkevi.ch/blog/#2026-09-27-vast-ai-releases-mira-scene-single-image-reconstruction-of</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-09-27-vast-ai-releases-mira-scene-single-image-reconstruction-of</guid>
			<pubDate>Sun, 27 Sep 2026 07:17:58 +0200</pubDate>
			<description>Mira-Scene reconstructs an editable 3D scene from one image and scores 0.727 3D-IoU on BlendSwap against 0.520 for SAM3D. Code and a checkpoint are public.</description>
		</item>
		<item>
			<title>Tencent Hunyuan and SJTU release AuK, a 1.5B open-weight speech editing model</title>
			<link>https://yurkevi.ch/blog/#2026-09-13-tencent-hunyuan-and-sjtu-release-auk-a-1-5b-open-weight</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-09-13-tencent-hunyuan-and-sjtu-release-auk-a-1-5b-open-weight</guid>
			<pubDate>Sun, 13 Sep 2026 10:38:42 +0200</pubDate>
			<description>AuK generates and edits speech from natural-language instructions, scores 2.65 average WER on Seed-TTS-Eval, and ships MIT-licensed with open weights.</description>
		</item>
		<item>
			<title>OpenAI claims finite-time blowup for forced Navier-Stokes</title>
			<link>https://yurkevi.ch/blog/#2026-09-12-openai-claims-finite-time-blowup-for-forced-navier-stokes</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-09-12-openai-claims-finite-time-blowup-for-forced-navier-stokes</guid>
			<pubDate>Sat, 12 Sep 2026 14:44:15 +0200</pubDate>
			<description>OpenAI says an internal model proved blowup for the forced Navier-Stokes equations on September 8, a day after Alpöge and Buckmaster posted the Euler case. Clay still lists the problem as open.</description>
		</item>
		<item>
			<title>Qwen releases Qwen3.8-Flash-Next, a 125B open-weight preview of Qwen4</title>
			<link>https://yurkevi.ch/blog/#2026-08-27-qwen-releases-qwen3-8-flash-next-a-125b-open-weight-preview</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-08-27-qwen-releases-qwen3-8-flash-next-a-125b-open-weight-preview</guid>
			<pubDate>Thu, 27 Aug 2026 00:14:07 +0200</pubDate>
			<description>Qwen3.8-Flash-Next is open-weight, has 125B parameters with 6B active per token, scores 62.5 on SWE-bench Pro, and previews the Qwen4 architecture.</description>
		</item>
		<item>
			<title>Z.ai confirms Ox Alpha is GLM-5.3-Flash, a 320B open-weight multimodal model</title>
			<link>https://yurkevi.ch/blog/#2026-08-26-z-ai-confirms-ox-alpha-is-glm-5-3-flash-a-320b-open-weight</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-08-26-z-ai-confirms-ox-alpha-is-glm-5-3-flash-a-320b-open-weight</guid>
			<pubDate>Wed, 26 Aug 2026 17:16:56 +0200</pubDate>
			<description>Z.ai revealed that the anonymous Ox Alpha model is GLM-5.3-Flash: 320B total, 18B active, MIT weights, 1M-token context and 84.3 on Terminal Bench 2.1.</description>
		</item>
		<item>
			<title>Tencent's Hunyuan3D-Buffalo 1.0 does text-to-3D, editing and understanding in one model</title>
			<link>https://yurkevi.ch/blog/#2026-08-10-tencent-s-hunyuan3d-buffalo-1-0-does-text-to-3d-editing-and</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-08-10-tencent-s-hunyuan3d-buffalo-1-0-does-text-to-3d-editing-and</guid>
			<pubDate>Mon, 10 Aug 2026 08:53:01 +0200</pubDate>
			<description>Tencent published Hunyuan3D-Buffalo 1.0 on August 3: text-to-3D, part-level editing and 3D understanding in one model, trained on 87M samples. Weights are not out.</description>
		</item>
		<item>
			<title>nyra labs releases CrisperWhisper 2.0, a 2B verbatim speech recognition model</title>
			<link>https://yurkevi.ch/blog/#2026-08-04-nyra-labs-releases-crisperwhisper-2-0-a-2b-verbatim-speech</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-08-04-nyra-labs-releases-crisperwhisper-2-0-a-2b-verbatim-speech</guid>
			<pubDate>Tue, 04 Aug 2026 09:21:52 +0200</pubDate>
			<description>CrisperWhisper 2.0 scores 87.8 disfluency F1 across 10 languages, ahead of ElevenLabs Scribe v2 and Microsoft, and the weights download and run offline for free.</description>
		</item>
		<item>
			<title>Eyeline Labs and Netflix release ID-V2V for identity-preserving video restyling</title>
			<link>https://yurkevi.ch/blog/#2026-08-04-eyeline-labs-and-netflix-release-id-v2v-for-identity</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-08-04-eyeline-labs-and-netflix-release-id-v2v-for-identity</guid>
			<pubDate>Tue, 04 Aug 2026 09:16:42 +0200</pubDate>
			<description>ID-V2V restyles a video's background and lighting from a single edited keyframe while holding faces, expressions and lip sync fixed. Weights and ComfyUI support are out.</description>
		</item>
		<item>
			<title>OpenAI's unreleased Astra model produces ten results on open math problems</title>
			<link>https://yurkevi.ch/blog/#2026-08-01-openai-s-unreleased-astra-model-produces-ten-results-on</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-08-01-openai-s-unreleased-astra-model-produces-ten-results-on</guid>
			<pubDate>Sat, 01 Aug 2026 18:45:31 +0200</pubDate>
			<description>OpenAI published ten proofs on long-open math and TCS problems today, produced by its unreleased Astra model, each shipped with a Lean 4 formalization.</description>
		</item>
		<item>
			<title>HKUST's HOMIE puts specific people and objects into generated video</title>
			<link>https://yurkevi.ch/blog/#2026-07-28-hkust-s-homie-puts-specific-people-and-objects-into</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-07-28-hkust-s-homie-puts-specific-people-and-objects-into</guid>
			<pubDate>Tue, 28 Jul 2026 09:01:09 +0200</pubDate>
			<description>HOMIE personalizes video with reference humans, products and logos, builds on the open Wan2.1-14B and Wan2.2-14B models, and ships code and weights.</description>
		</item>
		<item>
			<title>Microsoft releases Mage-Flow, a 4B image generation and editing model</title>
			<link>https://yurkevi.ch/blog/#2026-07-28-microsoft-releases-mage-flow-a-4b-image-generation-and</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-07-28-microsoft-releases-mage-flow-a-4b-image-generation-and</guid>
			<pubDate>Tue, 28 Jul 2026 08:55:49 +0200</pubDate>
			<description>Mage-Flow generates and edits images at native resolution, handles depth and segmentation, scores 0.90 on GenEval, and ships open under MIT.</description>
		</item>
		<item>
			<title>Anthropic releases Claude Opus 5, beating Fable 5 on most benchmarks at half the price</title>
			<link>https://yurkevi.ch/blog/#2026-07-24-anthropic-releases-claude-opus-5-beating-fable-5-on-most</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-07-24-anthropic-releases-claude-opus-5-beating-fable-5-on-most</guid>
			<pubDate>Fri, 24 Jul 2026 22:41:11 +0200</pubDate>
			<description>Claude Opus 5 shipped July 24, 2026. It outscores Fable 5 on most benchmarks at half the input price, though Fable stays ahead on SWE-bench Pro.</description>
		</item>
		<item>
			<title>Nvidia releases ARDY, an open real-time text-to-motion model</title>
			<link>https://yurkevi.ch/blog/#2026-07-20-nvidia-releases-ardy-an-open-real-time-text-to-motion-model</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-07-20-nvidia-releases-ardy-an-open-real-time-text-to-motion-model</guid>
			<pubDate>Mon, 20 Jul 2026 09:24:49 +0200</pubDate>
			<description>ARDY generates 3D human and humanoid motion from text in real time; code is Apache 2.0 and weights are open, which could help generate synthetic motion data.</description>
		</item>
		<item>
			<title>Claude Fable produces a counterexample to the Jacobian conjecture</title>
			<link>https://yurkevi.ch/blog/#2026-07-20-claude-fable-produces-a-counterexample-to-the-jacobian</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-07-20-claude-fable-produces-a-counterexample-to-the-jacobian</guid>
			<pubDate>Mon, 20 Jul 2026 09:05:53 +0200</pubDate>
			<description>On July 19, 2026, Levent Alpöge announced a Claude Fable counterexample to the Jacobian conjecture: a map with constant Jacobian determinant -2 and no inverse.</description>
		</item>
		<item>
			<title>Demis Hassabis proposes a FINRA-style body to regulate frontier AI</title>
			<link>https://yurkevi.ch/blog/#2026-07-18-demis-hassabis-proposes-a-finra-style-body-to-regulate</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-07-18-demis-hassabis-proposes-a-finra-style-body-to-regulate</guid>
			<pubDate>Sat, 18 Jul 2026 13:12:49 +0200</pubDate>
			<description>In a July 14 manifesto Google DeepMind CEO Demis Hassabis called for a US-led, industry-funded AI standards body that could vet frontier models before release and even pause the industry.</description>
		</item>
		<item>
			<title>Moonshot AI releases Kimi K3, a 2.8-trillion-parameter open-weight model</title>
			<link>https://yurkevi.ch/blog/#2026-07-18-moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open</link>
			<guid isPermaLink="true">https://yurkevi.ch/blog/#2026-07-18-moonshot-ai-releases-kimi-k3-a-2-8-trillion-parameter-open</guid>
			<pubDate>Sat, 18 Jul 2026 12:54:39 +0200</pubDate>
			<description>Kimi K3 is the largest open-weight model so far, scores near Fable 5 on the Intelligence Index, and tops the blind-voted Frontend Code arena.</description>
		</item>
	</channel>
</rss>
