EN RU

Daily digest

18 items · ~18 min · Week 2026-W35

Must-read (2)

Apodex 1.1: Scaling Agentic Intelligence for Complex Work

Apodex
Research official + media 2 src. ~1 min

Technical report defining 'working capability' — sustained, verifiable progress on long-horizon real-world tasks that touch files, sources, and code. Scales along two axes: environment scaling (file/search/code executable verifiers) and agentic coordination scaling (decomposition, async delegation, replanning). A shared AgentOS harness maintains task state and provenance across tools and agents; a 35B-parameter Apodex 1.1 Mini retains the capability locally.

Why it matters
HF Daily paper with 219 upvotes on Aug 25 — #1 paper on HF Daily that day. Frames a 'Heavy-Duty Solver' vision and shows smaller open model competitive with frontier systems on finance/research/math/coding/search benchmarks.
#agents #long-context #rlhf #agentic-rl #coding-agent #scaling #long-horizon

Alibaba launches Wan 3.0 video model with 30-second single-pass generation and native audio

Alibaba (Tongyi Lab)
Video official + media 8 src. ~2 min

Alibaba Cloud released Wan 3.0 on 2026-08-24 as a callable API on Model Studio / Qwen Cloud and the create.wan.video playground. The model generates up to 30 seconds in a single pass (double the 15-second ceiling of Wan 2.7), outputs native 1080p video with audio rendered in the same pass, and accepts text, image, video, audio and document (PDF / slide / web page) inputs. A 'thinking' mode is required for document inputs and lets the model reason about composition before rendering. Three endpoints ship at launch — T2V, I2V (with optional end-frame control), and reference-to-video — accepting up to 10 reference images, 5 reference video clips (15 s total at 16 fps+) and 5 reference audio tracks (15 s total). Pricing is $0.05/$0.10/$0.20 per second at 480p/720p/1080p, with a 'Prime' tier at $0.068/$0.14/$0.28. A public beta had been running since 2026-08-06; the Monday announcement coincided with the close of Alibaba's HK$80 billion follow-on share sale — the largest in Hong Kong since 2021.

Why it matters
Wan 3.0 doubles the single-pass video length ceiling set by the prior Wan generation and adds same-pass native audio, bringing Alibaba's flagship video model into the same feature bracket as ByteDance Seedance 2.5 and Hailuo H3. The reference composition (images / video clips / audio clips / documents as first-class inputs) positions it as a unified multimodal model rather than a pure text-to-video system. Wan remains API-only — Wan 2.2 is still the last open-weights flagship, so for open-source consumers this is not yet a drop-in replacement.
#video #text-to-video #image-to-video #audio #multimodal #alibaba #wan #tongyi #china #api #document-to-video #reference-to-video #video-generation

Worth knowing (3)

Claude Code v2.1.239 — major release with data-residency cost premium, cross-session messaging on Windows, ListAgents

Anthropic
Tools official 1 src. ~1 min

On Aug 21, Anthropic released Claude Code v2.1.239 (50+ changes): cost estimates (/cost, status line, --max-budget-usd) now include the 1.1x US-only-inference premium for data-residency workspaces; new /claude-api upgrade migrates Python projects from anthropic 0.x to 1.x; Alpine/musl builds now load native image-paste, clipboard, and audio-capture add-ons; cross-session messaging expanded to Windows (was macOS/Linux only); /goal check-ins back off (30 min -> 1 h -> 2 h) instead of firing every 30 min; ListAgents / /list-agents lists live teammates; keybindingFlavor 'readline' matches Bash word-key behavior. Also fixes Bedrock streaming billing, Edit/Write latency in JetBrains, MCP elicitation clipping, and OpenTelemetry trace fragmentation.

Why it matters
This is the largest Claude Code release in the window and the first to surface data-residency pricing in the CLI itself — a meaningful change for enterprise users on Bedrock/Vertex. The cross-session messaging expansion to Windows and the /goal backoff together signal Anthropic is investing in long-lived, multi-agent Claude Code workflows rather than single-session chats.
#anthropic #claude-code #cli #coding-agent #release #data-residency #bedrock #windows #agent-runtime

Anthropic Python SDK v1.0.0 — first stable major; httpx2 upgrade with breaking changes

Anthropic
Tools official 1 src. ~1 min

First stable major of the Anthropic Python SDK, released Aug 20. Upgrades to httpx2 with breaking changes documented in MIGRATION.md. Also stops warning about output_format= on the parse/stream/tool_runner helpers and restores original event imports in lib/streaming/_types.py.

Why it matters
Major-version bump of the official Python SDK is a maintenance signal for downstream projects pinned to 0.x — pinning will need to change. Worth flagging for anyone running production code against the SDK.
#anthropic #sdk #python #release #ga #developer-api

Anthropic TypeScript SDK v0.119.0 — Files and Skills APIs go GA; computer-use and browser-use toolsets

Anthropic
Tools official 1 src. ~1 min

TypeScript SDK release on Aug 19 (v0.119.0, mirrored as bedrock-sdk v0.33.0). Files and Skills APIs are now GA. Adds computer-use and browser-use toolsets. Followed on Aug 19 by v0.120.0 with managed-agents web search config and self-hosted sandbox memory. Multiple Bedrock/Vertex/Foundry/AWS SDKs shipped in lockstep.

Why it matters
Files + Skills going GA alongside new computer-use and browser-use toolsets is the API surface that backs Claude Code's MCP-connectors-with-managed-auth direction. For SDK consumers this is the canonical entry point for the new tools.
#anthropic #sdk #typescript #release #ga #computer-use #browser-use #managed-agents
For reference (13)

Mistral and HUMAIN announce strategic collaboration on sovereign AI in Saudi Arabia

Mistral
Industry official + media 3 src. ~1 min

On Aug 24, 2026, Mistral and HUMAIN announced a strategic collaboration worth hundreds of millions of euros to advance sovereign AI in Saudi Arabia and the broader Middle East. Initial focus areas are cybersecurity, voice technologies, frontier Arabic-language models, and regulated industries. Mistral will explore using HUMAIN's data-center infrastructure for local compute, with a joint go-to-market targeting regulated Saudi industries.

Why it matters
Marks Mistral's push into Gulf sovereign-AI deals, complementing its earlier European compute expansion with Microsoft; positions Arabic-language models as a frontier differentiator for non-US providers competing in regulated markets.
#mistral #humain #sovereign-ai #saudi-arabia #partnership #industry #compute #arabic #cybersecurity #voice-ai #regulated

Yandex opens medical AI assistant to all Russian doctors

Yandex
Industry official + media 6 src. ~1 min

Yandex opened beta access to its medical AI assistant (Yandex Med) to all practicing physicians in Russia, expanding beyond the previous partner-clinic-only pilot. The assistant transcribes doctor-patient dialogue into medical records, searches 1,000+ verified medical sources, and includes 20+ clinical calculators and severity scales (e.g. CURB-65). Pipeline uses a Classifier→Router→Reranker architecture for hallucination mitigation. Built on Yandex AI Studio, Yandex SpeechKit, and Yandex GPT. Available via web and Yandex Med Android app on RuStore; iOS version planned.

Why it matters
First Russia-wide rollout of a Russian-lab medical AI assistant to the full practicing-physician population, not just partner clinics. Builds on the August 4, 2026 spinoff of Yandex Med as a separate legal entity, signaling a dedicated healthcare-AI business line at Yandex.
#russia #yandex #yandexgpt #healthcare #product-launch #speech #yandex-med #speechkit #industry #enterprise

Prime Agent: A Self-Improving RLM Harness

Prime Intellect
Research official + media 3 src. ~1 min

Open-source harness for long-horizon coding and evaluation workflows. Persistent IPython REPL implements the Recursive Language Model abstraction; Continual Harness preserves histories/memories/skills/subagent specs across trajectories; recursive subagents coordinate via direct agent-to-agent messaging. On ARC-AGI-3 RHAE Best@1 the harness lifts a base model from 30% to 95.5%, and matches or beats native and popular harnesses on long-context coding, GPU kernel generation, emulator construction, and autonomous nanoGPT speedruns.

Why it matters
HF Daily paper with 18.2k cumulative upvotes — a single harness choice is enough to more than triple ARC-AGI-3 performance, suggesting many 'model capability' numbers are partly harness artifacts. Demonstrated refinement and parallelized subagents on Factorio.
#agents #rl #long-context #tool-use #harness #coding-agent #agent-runtime

ReWorld: An Interactive World Model with Long-Horizon Memory

TongyiLab (Alibaba)
Research official + media 2 src. ~1 min

Interactive video world model that separates short-horizon control from unbounded memory during training and bounds them at inference. Mixed per-head attention routes most heads to recent past while a few global heads attend over full history; a fixed-budget pose-indexed landmark KV cache keeps the entire past inside a constant memory budget; distribution-matching distillation in a LoRA adapter compresses sampling to 4 steps. Streams 704x1280 video in real time.

Why it matters
On 64-second out-and-back rollouts, a fixed 12-chunk cache regenerates the starting view where sliding windows lose evidence and full-KV attention OOMs. Best control fidelity (11.95 deg rotation error) and best generation quality against six recent world-model baselines.
#multimodal #long-context #world-models #video-generation #distillation #alibaba #tongyi

Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence

Research official + media 3 src. ~1 min

Position/survey paper by 35 authors across Jilin University, HKUST, NTU, Westlake and others. Argues single-agent paradigms (Prompt / Context / Harness / Loop Engineering) hit a structural ceiling on tasks needing heterogeneous expertise, parallel execution, independent verification, and persistent state. Introduces 'System Intelligence' and proposes 'Graph Engineering' — explicit, dynamic, evolving graph structures for tasks, agents, and runtime state, organized around Task Organization, Agent Coordination, and Runtime State Management views.

Why it matters
Paradigm-defining survey with an open awesome-list (DEEP-JLU/Awesome-Graph-Engineering). HF Daily listing on Aug 24 with 39 upvotes. Likely to anchor the next wave of multi-agent system papers.
#agents #multi-agent #systems #survey #icml-2026 #framework

InfinityEdit: Infinite Video Editing with a Lightweight Edit-Ignition Adapter

Alibaba (Tongyi)
Research official + media 2 src. ~1 min

Formalizes 'infinite video editing': given a preceding segment and an edit request, the model must generate the next segment that continues the stream under that edit, repeating indefinitely. A lightweight adapter with three attention blocks (history cross-attention, causal temporal self-attention, edit cross-attention) is selectively activated only in the chunk where an edit arrives; later chunks revert to the base model with a reset anchor.

Why it matters
Moves video editing from per-clip 'in-place' assumption to a streaming formulation, opening up long-form and live-editing workflows. HF Daily paper with 33 upvotes on Aug 24; 12-author Alibaba/Taobao team.
#multimodal #video-generation #video-editing #streaming #alibaba #tongyi #long-video

OpenCode v1.18.22 — OpenCode Go pricing cleanup, device-login fix, OpenAI-compat textVerbosity guard

SST
Tools official 1 src. ~1 min

Aug 24 release of SST OpenCode. Removes outdated OpenCode Go first-month discount messaging/pricing. Fixes device-login links when servers return relative verification URLs or use a base path. Stops sending textVerbosity to OpenAI-compatible providers that don't support it (community contribution by @joelstucki-taulia). Updates the Amazon Bedrock provider for compatibility fixes. Desktop: keeps model-provider headers visible while scrolling the model picker.

Why it matters
Cleans up the OpenCode Go pricing surface (the messaging was stale) and fixes two reliability issues — relative-URL device login and unsupported textVerbosity on OpenAI-compat endpoints — that would have broken setup on common provider configurations.
#sst #opencode #cli #coding-agent #release #openai-compatible #amazon-bedrock

Zed v1.16.2 — filesystem sandbox escape fix, GitHub Copilot Chat on GitHub Enterprise Cloud

Zed
Tools official 1 src. ~1 min

Zed stable release on Aug 24 (v1.16.2). Fixes a potential filesystem sandbox escape when running extensions (#63147); fixes GitHub Copilot Chat authentication and API routing for GitHub Enterprise Cloud (#63142); fixes project-level language server settings being ignored by extensions built with extension API versions up to v0.1.0 (#63083); fixes Flatpak CLI launches opening unrelated files due to an argument-construction bug (#62959). Pre-release v1.17.0-pre (Aug 19) added tabular CSV/TSV/PSV/SSV previews, new git blame/stash actions, and lower memory on large files.

Why it matters
The sandbox-escape fix is a security note worth flagging — anyone running untrusted extensions on prior 1.16.x should upgrade. The GHE Copilot Chat routing fix unblocks enterprise users who were hitting auth failures on Zed.
#zed #ide #coding-agent #release #security #github-copilot #sandbox #bug-fix

LangChain core 1.6.0 — standard model exception types, StructuredTool postponed annotations, RunnablePick deserialization

LangChain
Tools official 1 src. ~1 min

LangChain core 1.6.0 released Aug 19. Standard model exception types (adopted the same day by langchain-openai 1.6.0, langchain-anthropic 1.6.0, langchain-fireworks 1.6.0); StructuredTool postponed annotations; RunnablePick deserialization. langchain 1.3.16 (Aug 20) added standard model exception types and a custom token_counter in ContextEditingMiddleware. langchain-perplexity 1.4.1 (Aug 21) adds type='message' on Responses items, preserves extra_body, and bumps pillow. langchain 1.3.17 (Aug 25) adds custom HITL rejection reasons.

Why it matters
Standardizing model-exception types across providers is a real cross-package API quality win — code that catches the same exception class now works across OpenAI, Anthropic, and Fireworks without provider-specific branches. Worth flagging for anyone maintaining agents that span providers.
#langchain #framework #release #devtools #structured-output

LlamaIndex v0.14.24 — Claude Sonnet 5 / Opus 5 / GPT-5.6; gemini 3.7 Flash default; mcp 2.x

LlamaIndex
Tools official 1 src. ~1 min

LlamaIndex v0.14.24 released Aug 19. llama-index-llms-anthropic 0.11.10 adds Claude Sonnet 5, fixes Sonnet 5 function calling, adds Opus 5 to allowlists, corrects Opus 4.6 context window to 1M. llama-index-llms-bedrock-converse 0.14.18 mirrors Sonnet 5 + Opus 5 allowlists and adds thinking type 'disabled'. llama-index-llms-google-genai 0.10.0 defaults to gemini 3.7 Flash, fixes thought signatures. llama-index-llms-openai 0.7.10 adds GPT-5.6 models. llama-index-protocols-ag-ui 0.4.0 supports multimodal user input (image/audio/video/document). llama-index-tools-mcp 0.5.0 migrates to mcp 2.x.

Why it matters
Single release that brings Sonnet 5 / Opus 5 / GPT-5.6 / gemini 3.7 Flash into the LlamaIndex ecosystem at the same time. The mcp 2.x migration is a notable dependency bump for anyone using MCP tools via LlamaIndex.
#llamaindex #framework #claude #gpt-5.6 #gemini #release #mcp #model-catalog

llama.cpp b10603-b10615 — GLM-4.5-Air MTP, Deepseek 4 -sm tensor, Metal flash-attn vec tuning for M1 Pro/M2 Ultra/M5 Max

ggml-org
Tools official 1 src. ~1 min

Series of llama.cpp master builds Aug 23-24. Notable: b10603 adds MTP (multi-token prediction) support for GLM-4.5-Air; b10604 adds Deepseek 4 '-sm tensor' support; b10608 fixes mtmd-video moov atom at file end; b10610 shortens virtual-device naming in CUDA and Metal; b10614 splits per-op Metal sources for parallel compile and adds kernels for col2im_1d, set_rows, Q2_0, CONV_2D_DW, fused snake activation, FWHT; b10615 adds per-device tuned (Q, NE) flash-attn vec, including FA-vec tunings for M1 Pro, M2 Ultra, and M5 Max.

Why it matters
GLM-4.5-Air MTP and Deepseek 4 -sm tensor are model-specific perf paths that matter for anyone running those architectures locally. The Metal flash-attn vec tuning for M5 Max is forward-looking — it implies Apple silicon users on the newest chips will see meaningful speedups without changing code.
#llama-cpp #inference #metal #mlx #release #apple-silicon #glm #deepseek-v4 #performance

SGLang v0.5.18

SGLang
Tools official 1 src. ~1 min

SGLang v0.5.18 released Aug 22. Detailed changelog body was not surfaced via WebFetch — included for completeness; specific features unverified.

Why it matters
SGLang is one of the major open-source inference engines alongside vLLM and llama.cpp. A new release in this window is worth a heads-up for anyone tracking inference-engine versions, but the lack of surfaced detail means readers should consult the release notes directly before upgrading.
#sglang #inference #release #serving