DeepSeek V4 Flash, Orca + OMP, Suno Singing Ads — AI Daily Jul 31

572 messages · 79 active members

572
messages
79
active members
@mb29266, @jonmacofficial, @arielletolome
top contributors

Overview

DeepSeek dropped V4-Flash-0731 (ds4f) today and it dominated the conversation — 153M tokens for 79 cents, 200-250 TPS on dual RTX 6000s, 100-300 tok/sec on M5 Max 128GB running IQ2, and a 1M context window. GGUF quants hit Unsloth immediately. Consensus: unbeatable for high-throughput coding loops and copy generation, but GLM 5.2 still wins on heavy reasoning. Skeptics called it benchmaxxed while others said it's finally competitive with Opus 4.8. On tooling, Orca is emerging as the preferred UI wrapper — running OMP, OpenCode, and Codex harnesses side-by-side with native Linear integration and worktree assignment. One builder documented a full transition to Orca on VPS/Mac/iPhone with Hermes orchestrating Sol 5.6 and OMP running Grok 4.5. Meanwhile @jonmacofficial and others pushed Codex desktop as the harness winner over Hermes, Buzz, Claude, and Kimi, especially with OpenAI's 80% Luna and 20% Terra price cuts plus 2-3x Codex weekly resets making the $20 plan absurdly good value. Creative side: @mb29266 and @WhiskeyATX shared a Suno singing-ad masterclass (Claude writes lyrics → Suno generates song → animate), with GreenWave-style avatars crushing Meta at near-zero cost. Meituan open-sourced LongCat-Video-Avatar 1.5 for multi-minute talking-head video from one image plus audio. Recurring frustration: Claude/Fable/Opus 5 refusing legitimate work while Chinese models (GLM, Kimi, DeepSeek) stay permissive and win defections.

Topics

DeepSeek dropped V4-Flash-0731 with GGUF quants on Unsloth: 153M tokens for $0.79, 200-250 TPS on 2x RTX 6000, 100-300 tok/sec on M5 Max 128GB, and 1M context. Builders say it's unbeatable for volume coding/copy work but GLM 5.2 still wins heavy reasoning. Some call it benchmaxxed, others say it's finally catching Opus 4.8.

Orca is emerging as the preferred UI wrapper, running OMP, OpenCode, and Codex harnesses side-by-side with native Linear integration. Builders are pairing it with Hermes for orchestration and OhMyPi for coding, replacing Discord+Hermes and VSCode+cmux setups. Codex desktop with any model plugged in via CLI OAuth is also winning against Buzz and Claude.

@mb29266 and @WhiskeyATX shared how brands like GreenWave are turning ad scripts into songs via Suno then animating them for Meta. Workflow: Claude writes lyrics → Suno generates song → animate. Cost is minimal and organic engagement on VSL background music is spiking.

Meituan released LongCat-Video-Avatar 1.5, generating multi-minute realistic talking videos from a single image plus audio file. Requires proper GPU infrastructure but eliminates monthly avatar subscription costs and slots directly into n8n creative workflows.

OpenAI slashed GPT-5.6 Luna pricing by 80% and Terra by 20% while improving Sol's API speed. Codex is now dropping 2-3 weekly resets, effectively tripling the $20 plan's value. Meanwhile builders vent that Claude/Fable/Opus 5 feel lobotomized — ignoring memory files and refusing benign tasks — pushing defections to Sol and Chinese models.

Key Takeaways

  • DeepSeek V4-Flash-0731: 153M tokens for $0.79, 200-250 TPS on dual RTX 6000s, 1M context — practical local frontier-adjacent inference for high-volume coding and copy.
  • Orca is becoming the preferred UI to run OMP, OpenCode, and Codex harnesses side-by-side, with native Linear integration and worktree assignment killing older Discord+Hermes setups.
  • Suno + animation is a live ad format on Meta: Claude writes lyrics → Suno generates song → animate, scaling for pennies with organic engagement spikes.
  • LongCat-Video-Avatar 1.5 (open source) generates multi-minute talking-head videos from one image + audio — kills avatar subscriptions if you have GPU.
  • Codex weekly resets dropping 2-3x plus 80% Luna price cuts make cheap tiers viable for agent workflows; meanwhile Claude/Fable refusals are pushing builders to GLM, Kimi, and DeepSeek.

Hot Threads

@mb29266started

Suno singing ads and GreenWave avatar formats scaling on Meta

28 replies6 participants
@watchmedropshipstarted

DeepSeek V4-Flash-0731 release and local hardware benchmarks

20 replies7 participants
@fuckyesiwannatalkbusinessstarted

Full transition to Orca + Hermes + OhMyPi stack across VPS/Mac/iPhone

14 replies4 participants

Linked Items