Daily digest

9 items · ~9 min · Week 2026-W33

Worth knowing (3)

DeepSeek V4 Pro exits preview as V4-Pro-0813 with major coding benchmark gains

DeepSeek
Models / LLM official + media 3 src. ~1 min

DeepSeek quietly shipped DeepSeek-V4-Pro-0813 as the general-availability release of its V4 Pro flagship on August 12, 2026, ending a nearly four-month preview. The 1.6T-parameter MoE model (49B active) keeps its 1M-token context and shows large jumps in agentic coding benchmarks versus the preview, including SWE-bench Verified at 80.6%, LiveCodeBench at 93.5%, and Terminal-Bench 2.1 rising from 72.1 to 87.9.

Why it matters
It signals DeepSeek is closing the gap with top-tier closed models on agentic coding tasks while holding aggressive pricing, intensifying the China vs. US frontier-model price and capability race.

On-Policy Self-Distillation without Any Supervision

Research official 2 src. ~1 min

U-OPSD trains a language model on consensus pseudo-solutions built from its own majority-voted generations, correcting confident mistakes without ground-truth labels, teacher models, or environment feedback.

Why it matters
187 upvotes on HuggingFace Daily Papers; reports 8.5-10.7% gains over base models on math benchmarks (AIME24/25, HMMT25, MATH500, AMC23), competitive with supervised self-improvement methods.

Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA

Research official 2 src. ~1 min

An open agent-model system for post-deployment continual learning, combining a 744B GLM-5.2 base with a Mixture-of-LoRA architecture that composes specialist adapters for chat, agent, coding, and generative UI tasks; a smaller 50B Qwen3.6-based variant supports local deployment.

Why it matters
324 upvotes on HuggingFace Daily Papers; introduces a full co-designed stack (MinT post-training platform, LongStraw long-context RL, MindForge agentic RL) rather than a single isolated technique.
For reference (6)

Stealing Reasoning Traces from Proprietary LLM APIs

Research official 2 src. ~1 min

Researchers show that encrypted reasoning-trace blocks returned by major LLM APIs are interchangeable across sessions, users, and even different models, letting attackers inject them into weaker models to force verbatim disclosure of the plaintext reasoning.

Why it matters
Documents four attack vectors affecting Anthropic, OpenAI, and Google, recovering 367 PII artifacts and 182 credentials from over 315,000 scraped reasoning blocks; proposes cryptographic and system-level mitigations.

Yandex launches Lumena AI companion feature in Yandex Books

Yandex
Tools official + media 2 src. ~1 min

Yandex introduced the first capability of Lumena, a personalized AI companion planned across its entertainment services, starting with an 'About Characters' feature in Yandex Books that lets readers learn about characters and plot without spoilers, based on their own reading progress.

Why it matters
Signals Yandex's push toward personalized AI companions embedded across its consumer entertainment ecosystem (Books, Kinopoisk, Music), a new product surface beyond its core assistant Alice.

Yandex AI assistant helps pediatric rheumatologists remotely monitor juvenile arthritis patients

Yandex
Tools official + media 2 src. ~1 min

Yandex, built on Yandex AI Studio and trained on Russian Ministry of Health clinical guidelines and drug documentation, deployed an AI assistant inside a remote-monitoring app used by 32 pediatric rheumatologists tracking around 1,500 juvenile arthritis patients across Russia.

Why it matters
A concrete deployed healthcare use case for a Russian LLM platform, reportedly increasing the share of patient inquiries doctors can resolve on first contact by 40%.

Claude Code v2.1.229 adds self-hosted-runner hooks and plugin marketplace command sources

Anthropic
Tools official 1 src. ~1 min

Following v2.1.228's skill-sync hardening, Anthropic shipped Claude Code v2.1.229 on August 12, adding self-hosted-runner hook support, plugin marketplace 'command' sources that let a local tool supply and hot-reload plugin directories, VSCode sidebar session groups, and SSE keepalive for gateway streaming.

Why it matters
Coding-agent priority item: continues the skill-sync security hardening from the previous release with new CI-runner reliability and IDE workflow features for daily Claude Code users.

Zed editor 1.16.0 preview adds Gemini 3.6 Flash support

Zed Industries
Tools official 1 src. ~1 min

Zed's 1.16.0 preview release (Aug 12) adds Gemini 3.6 Flash to its Google AI model list, alongside Git panel improvements (collapsible grouped changes, optional stash messages) and zoomable/scrollable Mermaid diagram rendering.

Why it matters
Adds a newly released Gemini model to Zed's agent-capable editor, keeping its AI coding assistant current with the latest model lineup.

OpenCode v1.18.17 and v1.18.18 ship provider routing and reliability fixes

SST
Tools official 1 src. ~1 min

SST's OpenCode shipped v1.18.17 (Aug 12) and v1.18.18 (Aug 13), fixing the Kimi system prompt for Moonshot/Kimi providers, xhigh reasoning effort on xAI models, DeepSeek V4 Flash sampling defaults, session-compaction quality for smaller models, and enabling PDF vision attachments for GitHub Copilot models.

Why it matters
Coding-agent priority item: incremental reliability and model-routing fixes for OpenCode's multi-provider CLI/desktop app.