Kimi K3, GLM 5.2 Abliterated, Claude Resets — AI Daily Jul 16

669 messages · 81 active members

669
messages
81
active members
@geilt, @jcartu, @jarvisballer
top contributors

Overview

Today's chat centered on the arrival of Kimi K3 and its native CLI harness, which @geilt broke down in depth: AgentSwarm with 128 subagents (8 concurrent), 1M context window, HighSpeed mode (~5-6x output at 3x quota), durable overnight goals, and multi-provider support via config.toml — all at a ~$160/mo sub tier. Rumored at 3T params with claims of beating Fable 4.8, though the lack of published benchmarks drew skepticism. In parallel, @jcartu ran an overnight abliteration of GLM 5.2 on 8x H200s, producing a mostly-uncensored local model that keeps tool calling fully intact — the catch being you need 8x B200s or 16x RTX 6000 Pros for usable TPS. Simultaneous weekly resets across Claude, Codex, and Grok hit builders across timezones, layered on top of a midday Claude 529 outage and renewed Anthropic frustration over throttling, safeguards, and Fable feeling 'dumber' post-launch. The winning meta from @realcrischico: Grok 4.5 drafts fast, Fable reviews and patches, Grok executes, Fable does final review — with GPT-5.6 and Kimi increasingly filling in for Anthropic. The 'perfect 4' cited: Grok, Codex, GLM 5.2, and Kimi. On the stack side, @geilt detailed his bespoke Postgres/Valkey-driven fleet (container isolation, credential-blind tool calls, no MD files) as an alternative to Hermes and OpenClaw, while @jarvisballer explained Hermes as an operating layer routing Slack/Gmail/GitHub/Todoist into specialized agents with persistent memory. Orca emerged as a remote-VPS VS Code alternative. Other wins: @sav310 fired his accountant after AI invoicing uncovered months of billing errors in 2 days, and @aa12on shared a working multi-timeframe TradingView screenshot analysis system that's making LLM-driven trading viable again.

Topics

Kimi K3 dropped with a native CLI harness featuring AgentSwarm (128 total subagents, 8 concurrent), 1M context, HighSpeed mode, durable overnight goals, and multi-provider support via config.toml. Rumored 3T params with claims of beating Fable 4.8, though zero published benchmarks at launch drew skepticism. At ~$160/mo it's a serious contender in the harness wars.

@jcartu ran a 4-hour abliteration on 8x H200s to produce a mostly-uncensored GLM 5.2 with full tool calling intact — jumping on Tor, browsing forums, whatever you ask. The reality check: at 2.6-3T params, serving locally at usable TPS needs 8x B200s or 16x RTX 6000 Pros, making the $1500/yr GLM 5.2 unlimited coding plan more attractive than any homelab.

Weekly limits reset to zero simultaneously across Claude Max, Fable, and Codex, followed by a midday 529 outage on Claude. Frustration over throttling, aggressive safeguards, and Fable feeling 'lobotomized' post-launch has multiple builders (@tidemid, @geilt, @leewardbound) considering cancellations, with GPT-5.6 and Kimi cited as noticeably better for their workflows.

@jarvisballer detailed Hermes as an operating layer routing Slack/Gmail/GitHub/Todoist into specialized agents with persistent memory, tiered routing, and no-agent cron jobs — running cheap on Hetzner or Hostinger VPS. @geilt shared his bespoke Postgres/Valkey fleet (Obair, Triur, Naonúr) with container isolation and credential-blind tool calls. Orca emerged as a VS Code alternative that runs on a remote VPS. Multi-harness setups with synced global context, skills, and RAG'd sessions are the emerging pattern.

@sav310 is firing his accountant after an AI invoicing setup compressed 6 months of work into 2 days and uncovered duplicate conversion allocations and months of overbilling (Xero preferred over QuickBooks). @aa12on shared a working multi-timeframe TradingView screenshot analysis system for confluence-based entries, while @samb69 is reconsidering retired LLM trading setups given newer models — floating Cerebras Gemma 4B at 1800 tps for real-time tick analysis.

Key Takeaways

  • Kimi K3's native AgentSwarm (128 subagents, 8 concurrent) and 1M context at ~$160/mo make it a serious harness contender — but treat 'beats 4.8' claims as unverified until benchmarks land.
  • Locally-served frontier open weights (GLM 5.2, Kimi K3) are technically doable but economically absurd — a $1500/yr unlimited GLM plan beats any homelab you'd build.
  • Builders are actively rotating away from Anthropic due to throttling, safeguards, and Fable's post-launch quality drop — the 'perfect 4' cited: Grok, Codex, GLM 5.2, and Kimi.
  • Hermes' value isn't better code than CC — it's the operating layer routing Slack/email/GitHub into specialized agents with persistent memory and cheap tiered routing.
  • AI accounting/invoicing and multi-timeframe LLM chart analysis are genuinely underpriced automation opportunities right now — newer models make previously-failed setups viable again.

Hot Threads

@geiltstarted

Kimi K3 harness deep dive — AgentSwarm, HighSpeed mode, and multi-provider config

22 replies6 participants
@jcartustarted

Abliterated GLM 5.2 with tool calling intact — and what it takes to serve locally

28 replies6 participants
@realcrischicostarted

Grok 4.5 + Fable + GPT-5.6 planner/executor workflow and Hermes setups

14 replies7 participants

Linked Items