A lightweight configuration circulating among developers lets Claude Code users flip between Anthropic's Fable 5, Zhipu AI's GLM and Moonshot AI's newly released Kimi K3 simply by typing claude, glm or kimi at the terminal — no external tooling required.
July 2026 · Kimi K3
One Terminal, Three Frontier Models — Just Type claude , glm or kimi
A lightweight settings file lets Claude Code users swap between Anthropic's Fable 5, Zhipu's GLM and Moonshot's new Kimi K3 with a single word — no proxy, plugin or external tooling required.
2.8T
parameters, open-weight MoE model
16/896
experts activated per token
1M
token context window, native vision
Frontend Code Arena — head-to-head
Higher is better · scores vary by evaluation harness
Kimi K3 also scored 88.3 on Terminal-Bench 2.1 (KimiCode harness).
How the switch works
settings/kimi-k3.json
base URL + API key + model name
→
shell alias
kimi() { claude --settings … }
→
one-word switch
routes Claude Code to any backend
Leans on Claude Code's support for Anthropic-compatible endpoints — the same pattern that existed in the Kimi K2 era, now sharpened by K3's arrival.
Early enthusiasm
Strong on front-end and vision-in-the-loop work — code generation, debugging, whole-repo understanding — with favorable cost-performance across game dev, chip design and research code.
Caution
Fable 5 leads slightly on overall intelligence and, tuned for Claude Code, can be more reliable over long sessions and less prone to hallucination. Read head-to-head numbers with care.
$3 / M
input tokens (cache hits lower)
Full weights due July 27 . As labs converge on Anthropic-compatible APIs, model choice becomes a per-command decision — not a locked-in dependency.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…