Opus 5 vs Fable, Hermes Stack, Kimi K3 in C — AI Daily Aug 2

321 messages · 63 active members

321

messages

active members

@jonmacofficial, @jcartu, @momodxb

top contributors

Overview

Model preferences shifted noticeably today — Opus 5 is quietly replacing Fable as the orchestration model of choice for several builders, with @fmill1 and @jonmacofficial reporting it beats Fable on usage limits and general reliability. @jcartu called Fable 'benchmaxxed' post-nerf and positioned Opus 5 as a solid ds4f upgrade for orchestrating Hermes stacks. Grok 4.5 got praise for front-end design, while Codex 5.6 Sol was called out for better visual polish than Claude CLI out of the box. A common top-three lineup emerged: Opus 5, GLM 5.2, and Kimi K3 — the latter now runnable CPU-only in 8.24 GB of RAM via a portable C99 implementation. Alibaba also dropped Qwen 3.8, their most capable model yet. On the Hermes stack, members compared personal forks and flagged Telegram's rate-limit friction under heavy load; direct SSH or web admin was recommended for serious work, while mobile builders defended Telegram for turn/memory visibility. @jcartu also shared advanced context tips using Hindsight 0.8.6+ and OMP's /shake command to effectively extend context to 500k+ by dumping tool calls. Tooling saw a YC-backed agent harness (quantum machine), Macro's all-in-one workspace launch, and Penpot MCP pitched as a much better open-source alternative to Figma's MCP. VoxCPM2 emerged as the go-to open-source ElevenLabs replacement, runnable on 8GB VRAM. LATAM hiring sparked substantive debate, with mixed results and Paired.so floated as a higher-tier alternative to Somewhere.

Topics

Opus 5 Overtakes Fable for Orchestration22 msgs

Multiple builders reported switching from Fable to Opus 5 for orchestration, citing better reliability and usage economics. @jcartu called Fable 'benchmaxxed' and positioned Opus 5 as a strong ds4f upgrade for Hermes stacks. Fable also came up as a Figma replacement candidate, though heavy token consumption is a real cost consideration. Grok 4.5 got a nod for front-end design tasks.

Hermes Stack: Customization and Telegram Friction10 msgs

Members compared personal Hermes forks referencing hermes and takopi. Telegram was praised for turn/memory visibility on mobile but criticized for rate-limit chaos under heavy load — direct SSH or web admin was recommended for serious use.

Model Lineup: Opus 5, GLM 5.2, Kimi K3, Qwen 3.85 msgs

A quick-fire share surfaced Opus 5, GLM 5.2, and Kimi K3 as the current go-to trio. Kimi K3's 2.78T parameters now runs CPU-only inference in 8.24 GB via portable C99 — no BLAS, no GPU. Alibaba also unveiled Qwen 3.8, positioned as their most capable model yet and close behind Moonshot in size.

Hindsight + OMP Context Management Tricks12 msgs

@jcartu detailed how Hindsight 0.8.6+ combined with OMP's /shake command can dump 20-30% of context (tool calls, no real content) to effectively give 500k+ context windows. Users were advised to have their agents pull the latest Hindsight docs to ensure all optimizations are enabled on their banks.

VoxCPM2 as Open-Source ElevenLabs Alternative9 msgs

With ElevenLabs accounts continuing to get shut down, @jonmacofficial recommended VoxCPM2 for local voice cloning. Minimum requirements are modest: 8GB VRAM (RTX card), 16GB RAM, and 15-25GB storage. Fish Audio was also shared as a hosted alternative with an official LLM-friendly docs endpoint.

Key Takeaways

Opus 5 is now beating Fable for orchestration in real-world usage — Fable appears benchmaxxed post-release nerf.
Kimi K3 (2.78T params) can now run CPU-only inference in 8.24 GB via a portable C99 implementation.
Hindsight 0.8.6+ with OMP's /shake command effectively extends context to 500k+ by dumping tool call history.
Telegram is convenient for mobile Hermes visibility but breaks under heavy load — use SSH or web admin instead.
VoxCPM2 runs on 8GB VRAM for local voice cloning — solid ElevenLabs replacement as accounts get banned.

Hot Threads

@yangthegoatstarted

Paired.so vs Somewhere for LATAM hiring

14 replies7 participants

@taszacharisstarted

One-shotting Shopify landing pages from Figma designs

8 replies5 participants

@jaypozostarted

Hermes on Telegram: visibility vs rate limits

7 replies2 participants

Linked Items

Overview

Topics

Key Takeaways

Hot Threads

Linked Items

Scalability School - YouTube

Instagram

GitHub - mikefutia/ai-visibility-audit · GitHub

Buzz — Your people, your agents, your project — all in one place.

Jordan Menard on X