Daily digest
6 items · ~6 min · Week 2026-W35
Worth knowing (2)
DeepSeek releases experimental multimodal V4-Flash-Vision-Exp
DeepSeekDeepSeek shipped V4-Flash-Vision-Exp on its API platform as its first multimodal entry, matching V4-Flash on text tasks (agents, reasoning, world knowledge) while adding image understanding billed at up to 384 tokens per image. The vision model lands close to Opus-4.8 on DeepSeek's multimodal-agent benchmark table and ships with a free Files API for image reuse and DeepSeek Harness 0.1.1 day-one support.
OpenAI Codex CLI v0.149.1 stable adds thread-source classification and image-budget compaction
OpenAIOpenAI released Codex CLI v0.149.1 as Latest stable on 2026-08-24, alongside alpha tags 0.149.0-alpha.7.2 / 0.149.0-alpha.4.3 (22-23 Aug) and 0.150.0-alpha.7 (22 Aug). v0.149.1 introduces a global `codex exec --thread-source <SOURCE>` flag propagated to new and forked threads and surfaced as `threadSource` in the TypeScript SDK, an opt-in `compaction_image_budget` feature that charges retained images using existing size estimates with atomic image-and-label truncation, and sets `thread_source=memory_consolidation` for detached memory requests.
For reference (4)
ForgeWM: Progressive Causal Training for Few-Step Action-Conditioned Video World Models
Four-stage progressive training (domain adaptation -> teacher-forced causal -> causal consistency distillation -> on-policy distribution matching) converts a bidirectional video generator into a 1/2/4-step causal world model. Demonstrated on Minecraft and gamepad-controlled FPS with dual-path deployment for interactive vs replay use.
SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?
OpenMOSSRepository-level benchmark of 119 tasks across 98 GitHub repos in 20 scientific domains, spanning issue-driven, expert-exploratory and engineering-integration paradigms. Best agent (Claude Code on Opus) scores under 50% pass@1; the paper catalogues four recurring failure mechanisms.
Cline ships v4.1.13/14/15 with model catalog refresh, MCP auto-approve fix and `cline hub` drain/upgrade commands
ClineCline released v4.1.13, v4.1.14 and v4.1.15 across 2026-08-22 to 2026-08-23, plus matching CLI v3.0.57, SDK v0.0.78 and Desktop v0.0.16 drops. v4.1.14 refreshes the built-in model catalog with Claude Fable 5, Grok 4.6 on Vertex, DeepSeek V4 Flash variants (including a vision preview), MiMo v2.5, Qwen3.8 27B, Gemma 4 26B, LongCat 2.0, Nemotron 3.5 Lightning and Thinking Machines Inkling, and restores task-completion telemetry. v4.1.15 fixes MCP auto-approve so the 'Use MCP servers' toggle governs all MCP tools. CLI v3.0.57 adds `cline hub drain` and `cline hub upgrade`; the SDK and Desktop drops add durable hub handoff with event replay and dedup and Langfuse session/client identity.
OpenClaw 2026.8.1-beta.3 adds GPT-5.6 reasoning, CDP relay and verified Gateway restart handoff
OpenClawOpenClaw published pre-release 2026.8.1-beta.3 on 2026-08-24. It adds GPT-5.6 Sol, Terra, Luna and Ultra reasoning support across OpenClaw and the Codex runtime, a Puppeteer-compatible CDP relay for paired Chrome sessions, explicit external Gateway lifecycle supervision with verified restart handoff, compact verified SQLite backup and fresh-target restore commands, and shared durable ingress monitors for channel plugins. The bundled Codex runtime pins @openai/codex@0.149.1.