GLM 5.3 Flash at 320tps, OMP Harness, Video Model Shootout — AI Daily Aug 27

382 messages · 68 active members

382
messages
68
active members
@jcartu, @jonmacofficial, @mb29266
top contributors

Overview

Wednesday's builder chatter split between a landmark performance milestone and a wave of practical stack blueprints. @jcartu cracked GLM 5.3 Flash on four 6k's, hitting 320 tokens/sec with 2.3s TTFT and 12k prefill — reportedly the fastest known deployment, running on as few as two 6k's while beating GLM 5.2 at half the size. Alongside that, he flagged that Hugging Face's reported $13B valuation makes the concentration of open-model distribution risky, and he's cold-storing BF-16 weights on a 22TB NAS as insurance. On the harness side, @thewildzeno, @jcartu, and Kieran walked through why OMP is winning: model-agnostic swapping mid-session without losing context, LSP-driven background code fixing, and an auth broker for multi-machine setups. @jonmacofficial published a full production blueprint layering Orca as terminal manager, OMP as harness, Grok 4.6 as main agent, with Fable 5 for planning and GPT-5.6 Sol for large or security-sensitive PRs — plus hard rules like 240-char output caps and <800 lines per file to survive parallel worktrees. Creative-side, @jonmacofficial and @arielletolome ranked video models (Seedance 2.5 > 2.0 > Mini > Google Omni, with Kling 3.0 Turbo as the value pick), while @jasonakatiff's F12 discovery kicked off a thread on using browser Network tabs to reverse engineer undocumented APIs. VPS + Tailscale continues to emerge as the default for persistent agent boxes.

Topics

@jcartu reports GLM 5.3 Flash running at 320 tok/sec with 2.3s TTFT and 12k prefill on four 6k GPUs — claimed as world's fastest. It runs on just two 6k's, outperforms GLM 5.2 at half the size, and supports full precision at half the throughput. Builders compared output quality to Opus 4.8-class.

Extended discussion on why OMP beats Claude Code and OpenCode: model-agnostic design, LSP auto-fixing, mid-session model switching, and the auth broker for multi-machine setups. @thewildzeno shared custom mods including orchestrator modes with Gemini Flash, batch tool calls, and subagent speed monitoring. Pairs well with Paseo for mobile workflows.

@jonmacofficial detailed his production stack: Orca terminal manager, OMP harness, Grok 4.6 main agent, with role-split subagents (Fable 5 planning, Grok coding/fast review, GPT-5.6 Sol for PRs >2k lines or security). Key knobs: /vibe and /advisor modes, 240-char output caps, Ponytail plugin in Ultra mode, and enforcing <800 lines per file to avoid catastrophic merge conflicts in parallel worktrees.

@jonmacofficial and @arielletolome ran head-to-head comparisons for Viral View. Ranking: Seedance 2.5 > Seedance 2.0 > Seedance 2.0 Mini > Google Omni, with Kling 3.0 Turbo flagged as best price-to-performance for UGC. On images, GPT Image 2 and Qwen were community favorites, though Grok Imagine 2 won one prompt-adherence test. @boostimo noted the arena leaderboard shows fierce but uneven reliability.

@jasonakatiff's F12 discovery sparked a thread on browser Network tab tricks for reverse engineering undocumented APIs (request URLs, cookies, tokens). VPS + Tailscale is converging as the default for persistent agent boxes. @sibunting shared an overnight workflow: Fable drafts specs, Hermes on Sol runs adversarial review, then agents build overnight with Opus + Sol QA — resolving ambiguity at spec time, not runtime.

Key Takeaways

  • GLM 5.3 Flash hits 320 tok/sec on four 6k GPUs (2.3s TTFT) — runs on just two 6k's and beats GLM 5.2 at half the size.
  • OMP's edge is model-agnosticism — swap models mid-session without losing context, with LSPs continuously fixing code in the background.
  • Enforce <800 lines per file and 240-char agent output caps to avoid merge conflicts and reduce cognitive load in parallel worktrees.
  • Seedance 2.5 wins on quality but Kling 3.0 Turbo and Seedance 2.0 Mini are the value picks for high-volume UGC video.
  • Run agents on a VPS with Tailscale for persistent sessions, and cold-store open model weights — HF/ModelScope are the only distribution game in town.

Hot Threads

@jonmacofficialstarted

Video model comparison for Viral View and Orca+OMP+Grok stack breakdown

32 replies9 participants
@thewildzenostarted

OMP harness mods, orchestrator modes, and extensions

18 replies6 participants
@jasonakatiffstarted

F12 dev tools discovery and API reverse engineering tips

11 replies7 participants

Linked Items