Kimi K3 on Fireworks, Codex+Fable, Orca+Tailscale — AI Daily Jul 26

426 messages · 72 active members

426
messages
72
active members
@jasonakatiff, @Kieran, @jcartu
top contributors

Overview

Kimi K3 dominated the morning as builders hunted for non-China hosting options. @jcartu confirmed Fireworks.ai now hosts K3 in the US with near-lossless MXFP4 quantization (invite-only firepass initially, broader access expected the next day), removing the main data-sovereignty blocker for American builders. @jasonakatiff stress-tested K3's /swarms feature on Max — burning ~50% of a 5-hour quota in 15 minutes on a full-agent codebase review that included API-key leakage testing — sparking debate about whether swarm workflows and tight usage caps can coexist. K3 also drew praise as a generational leap for creative work and pen testing, though most builders still won't hand it their main proprietary codebase. Multi-model orchestration was the deeper thread. Rather than picking a winner between Codex and Fable, builders are running them in parallel: Codex reliably flags issues Fable misses on audits, while Fable still one-shots generation better. Basant paired Codex with Opus 4.8 for launch plans and competitive scraping. Post-relaunch Fable regressions pushed @csoares12 and @rmktg to unwind unwanted architectural additions, cementing consensus that Fable now belongs on frontend (paired with Mobbin/Higgsfield MCP) while backend and security work routes to Opus, K3, or Sol. @fuckyesiwannatalkbusiness shared a full stack: Sol 5.6 orchestrator, Grok 4.5 coder, Terra rescue workers, Gemini 3.6 Flash for 1M-context compression. On infrastructure and harness discipline, @sav310 detailed a canary architecture using Goliath as a persistent orchestrator that provisions fresh Hetzner CCX VMs per agent run and destroys them after — sandbox isolation, not burst capacity, being the point. @realcrischico and @thewildzeno endorsed Orca (onorca.dev) as a cross-platform cmux replacement, with Orca + Tailscale enabling phone-based control of multiple concurrent sessions. Kieran built a hybrid harness combining superpowers with Matt Pocock's skills library, while @jaypozo is trimming claude.md, pruning MCP tools, and swapping in mem0 as usage caps force tighter context discipline.

Topics

@jcartu confirmed Fireworks.ai now hosts Kimi K3 in the US with MXFP4 quantization described as nearly lossless and likely faster than Moonshot. Access is invite-only via firepass initially, with broader rollout expected the next day. K3 also drew praise as a generational leap for creative work and pen testing, though builders remain cautious about sending proprietary code to Chinese-hosted inference. @jarvisballer noted K3 with Codex as orchestrator works well thanks to Codex's free refreshes.

Multiple builders confirmed Codex catches blind spots Fable misses during audits, while Fable still edges out on one-shot generation — the emerging pattern is running both and merging outputs. Basant paired Codex with Opus 4.8 for launch plans and competitive feature scraping. Post-relaunch Fable regressions pushed builders to keep it on frontend work (with Mobbin and Higgsfield MCP) and route backend/security tasks to Opus, K3, or Sol.

@realcrischico and @thewildzeno endorsed Orca (onorca.dev) as a cross-platform cmux replacement available on Mac and Windows (WSL required). @realcrischico outlined an Orca + Tailscale setup: install Tailscale on both desktop and phone, pair via the tailnet, and remotely manage active CLI sessions. Claude's new Remote Control feature is an alternative but requires per-session initiation, whereas Tailscale gives blanket access.

@sav310 shared a canary architecture where Goliath acts as a persistent orchestrator that provisions fresh Hetzner CCX VMs per agent run, bootstraps scoped access, collects evidence, then destroys the VM. The sandbox boundary — not just burst capacity — is the primary goal. @jaypozo plans to borrow the pattern for scaling parallel loops.

Kieran built a hybrid harness combining the superpowers repo with Matt Pocock's skills library (github.com/mattpocock/skills), swapping in grillme and running /review with a GPT Sol sub-agent for cross-model review. @jaypozo is trimming claude.md, pruning MCP tools, and swapping in mem0 borrowed from Hermes. @seekersight shared a confidence-audit skill that fires before merging to main, labeling every claim CERTAIN/HIGH/MEDIUM/LOW/GUESS — the bigger win being that the audit pass forces the model to catch weak assumptions before labeling.

Key Takeaways

  • Kimi K3 is now hostable in the US via Fireworks.ai, removing the main data-sovereignty blocker — but keep proprietary codebases off Chinese-hosted inference.
  • Don't pick between Codex and Fable — run both in parallel and merge outputs; Codex catches audit blind spots, Fable one-shots better.
  • Orca + Tailscale beats Claude's per-session Remote Control for phone-based management of multiple concurrent agent sessions.
  • Ephemeral per-task VMs (e.g., Hetzner CCX) give real agent isolation, not just burst capacity — worth the provisioning overhead.
  • As plan caps tighten, trim claude.md, prune MCP tools, move repeated actions into granular skills, and force models to audit their own confidence before merging.

Hot Threads

@jasonakatiffstarted

Kimi /swarms burning through Max quota on codebase review

22 replies8 participants
Kieranstarted

Hybrid OMP harness with superpowers + Matt Pocock skills

14 replies6 participants
@sav310started

Ephemeral Hetzner worker architecture with Goliath orchestrator

12 replies3 participants

Linked Items