Kimi K3.1 Pro, Dessa Voice, Argon vs Grok 4.7 — AI Daily Oct 23
490 messages · 69 active members
Overview
Topics
Moonshot released Kimi K3.1 Pro with open weights promised Nov 4, reportedly timed ahead of possible Chinese restrictions. @jcartu says it's strong at ~60 tok/s but too large for most local rigs. The $200 subscription is disappointing — members burned 8.5% of the 5-hour quota in 10 turns on a 50k chat, so most are keeping it as a Claude fallback rather than a daily driver.
Block released Dessa, a conversation model trained only on speech, and a demo persona nicknamed 'Darius Johnson' had the whole group dying laughing. Builders are already imagining outbound calling use cases and next-gen features like muffled-mic interjections, with speculation on how far behind ElevenLabs now is. Access is currently gated behind a waitlist.
@rmktg's median latency tests: Grok 4.7 won on speed (10.4s tool task), Opus 5.5 second, Gemini Flash edged Argon on agent work. @tidemid prefers Argon for its grounded tone and lower refusal rate but admits it's weak at code — pair it with Codex. Argon nearly one-shot a 7-hour task where DeepSeek got stuck at stage 2.
Consensus Hermes stack: Opus 5.5 (medium) orchestrator, Grok for fast research/ripping, Sonnet/Opus for creative profiles, DeepSeek 1017 or GLM for grunt work. @iammikevineyard applied the same pattern to game dev — Opus 5.5 calling Codex as sub-agent, plus ElevenLabs, Tripo, Fal, Godot, Blender, and Claude Eyes in the loop — claiming ~15k hours of equivalent work in 30 days.
@leewardbound shared his Hetzner auction baremetal box (256GB RAM, ~$150/mo) running k3s with ~80 projects. Consensus: homelab for personal workloads and learning, rented/managed for business. @nowwatchthisdrive's hypothetical $50k budget kicked off Mac Studio vs dual-GPU debate, with RTX 5090 retail ballooning from $1999 to $8500 making unified-memory Macs increasingly attractive.
Words worth knowing
Three terms from the glossary.
- Vibe CodingWriting code by describing what you want in natural language without deep technical planning.
- Open ClawAn open-source version or alternative to Claude Code that members run locally on their own hardware.
- Claude CodeAnthropic's coding interface that uses Claude models to write and edit code with a planning system.
Key Takeaways
- Kimi K3.1 Pro gets open weights Nov 4 and is strong, but the $200 tier burns 5-hour limits too fast to replace Claude — keep it as a fallback.
- Dessa's speech-only training produces persona realism that leapfrogs current TTS — watch for outbound calling disruption.
- Argon 4 is the sleeper for long-horizon tasks (grounded, refuses less, one-shots work where DeepSeek stalls) but offload coding to Codex.
- Best Hermes stack right now: Opus 5.5 (medium) orchestrator, Grok for research/ripping, Sonnet/Opus for creative, DeepSeek 1017 or GLM for grunt work.
- Hetzner auctions yield 256GB RAM baremetal for ~$150/mo with k3s; meanwhile RTX 5090 retail jumped $1999 → $8500, making Mac Studio competitive for local models.
Hot Threads
Dessa 'Darius Johnson' voice demo and ElevenLabs implications
AI game dev pipeline — Opus + Codex sub-agents + Godot/Blender stack
Grok vs Opus vs Gemini latency benchmarks on Hermes tool tasks