EN RU

Daily digest

7 items · ~7 min · Week 2026-W31

Must-read (1)

MiniMax Launches H3, an Omni-Modal Model Generating 2K Video With Native Stereo Audio

MiniMax
Video official + media 4 src. ~1 min

MiniMax officially launched H3 on July 31, an omni-modal generation model that jointly understands text, image, video, and audio context to produce up to 15-second 2K video clips with native stereo sound, accepting up to 9 reference images, 3 video clips, and 3 audio clips as controllable inputs. The model is live via MiniMax's API and the Hailuo AI consumer app, with open weights promised in the coming days.

Why it matters
Native audio-video co-generation at commercial 2K quality, with open weights planned, pushes open Chinese labs (competing with ByteDance and Kuaishou) further into a segment previously led by closed Western video models, while undercutting rival pricing by roughly two-thirds.
#minimax #video-generation #omni-modal #audio #text-to-video #image-to-video #open-weights

Worth knowing (4)

DeepSeek Puts V4-Flash-0731 API Into Public Beta, Beating Its Own Flagship on Agent Benchmarks

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

DeepSeek released the V4-Flash-0731 build of its API into public beta on July 31, keeping the same architecture as the V4-Flash preview but re-post-trained to surpass DeepSeek's own V4-Pro-Preview on all nine published agent and coding benchmarks, including Terminal-Bench 2.1 and DeepSWE. The model natively supports the Responses API format and is compatible with Codex.

Why it matters
A cheaper 'flash' tier model now beats DeepSeek's flagship preview on agentic coding benchmarks, showing post-training alone can close much of the gap to a larger model.
#deepseek #v4-flash #agentic-coding #api #benchmark

Metis: Memory Foundation Model

MemTensor, Renmin University, NUS, Shanghai Jiao Tong University, Tongji University
Research official 1 src. ~1 min

Metis introduces a 'memory foundation model' that gives transformers a persistent, dynamically evolving native memory state via a Metis Block (Local Memory Block plus Hyper Memory Block) inserted into transformer layers, avoiding linearly growing context windows.

Why it matters
Reached 255 upvotes on HuggingFace Daily Papers; proposes an architectural alternative to external-memory/RAG approaches for long-term agent memory, with code open-sourced.
#memory #architecture #long-context #research

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Horizon Research, Frontis.AI, Tsinghua University
Research official 1 src. ~1 min

Frontis-MA1 (35B) is a meta-evolution agent post-trained around four program-evolution operators (Draft, Improve, Debug, Crossover) inside OpenMLE, an open full-stack recursive self-improvement testbed for machine learning engineering with verifiable execution feedback.

Why it matters
Gathered 165 upvotes on HuggingFace Daily Papers; on MLE-Bench Lite it lifts Medal Average from 39.39% to 60.61% (71.21% with extended search) over its base model, with full model weights and stack released.
#self-improvement #rl #agents #reasoning #research

PhiZero: A World Model Built Around Physical Language

NLPR, Institute of Automation, Chinese Academy of Sciences
Research official 1 src. ~1 min

PhiZero learns a compact discrete 'physical language' representing world-state transitions from unlabeled in-the-wild video, then uses a reason-then-render paradigm to infer future world evolution as a physical-language sequence before rendering it into video.

Why it matters
Collected 157 upvotes on HuggingFace Daily Papers; demonstrates zero-shot motion transfer and fine-grained action-conditioned simulation, an approach to world modeling distinct from pixel-space video diffusion.
#simulation #video-generation #research
For reference (2)

GitHub Copilot CLI ships v1.0.77-v1.0.78 with sandbox and session workflow improvements

GitHub
Tools official 2 src. ~1 min

Copilot CLI v1.0.77 (Jul 30) added a web-based OAuth login as default for local terminals, Ctrl+G editor-based answer editing, and MDM-enforced sandbox policy; v1.0.78 (Jul 31-Aug 1) sped up resuming long sessions (a 230MB/74k-event transcript loads in under a second at roughly a quarter of the previous peak memory), added a /permissions command, and fixed extension slash-command double-invocation and sessionEnd hook firing for piped runs.

Why it matters
Continues GitHub's push to make Copilot CLI a serious terminal coding agent competitor to Claude Code and Codex CLI, with session-resume performance work directly targeting long-running agent workflows.
#github-copilot #coding-agent #cli #mcp

OpenCode v1.18.11 fixes MCP reconnect loops and desktop stability issues

SST
Tools official 1 src. ~1 min

OpenCode v1.18.11 (Aug 1, 2026) stops MCP SSE connections from getting stuck in reconnect loops after server error responses, fixes provider configs using interleaved reasoning fields like reasoning_text, and improves desktop app stability (external links open in system browser, fixed stale session tab state, prevented file tree tab clipping).

Why it matters
Incremental reliability fixes for one of the main open-source Claude Code/Codex alternatives, particularly around MCP server connection stability which affects tool-use reliability.
#opencode #coding-agent #mcp #desktop #bug-fix