Skip to content

Stacks/ chinese-coding-stack

Chinese Coding Stack

Frontier-class agentic coding at a fraction of US subscription prices — Kimi, GLM, Qwen and DeepSeek as the daily drivers.

Open in Noemium

observed by @whysanesandersVerified 2026-08-19

$47/mo

difficulty: intermediate

Shoestring tradeoff. GLM Coding Plan Lite plus free Kimi/Qwen CLIs as backup. No Kimi hosted membership, no DeepSeek off-peak bulk. Daily small fixes stay cheap; long-horizon K3 sessions and batch test generation go away.

The components — what each one costs

  • Kimi CodeCLI is free/MIT; hosted coding membership ~$19–$39/mo depending on region; API billed separately at Kimi API rates (K3 $3/$15 per 1M)
  • GLM Coding PlanLite ~$18/mo, Pro ~$72/mo, Max ~$160/mo (annual billing $151.20/$604.80/$1,344/year); rate/allowance caps apply
  • DeepSeek APIno free API tier; V4 Flash peak $0.44 input (cache miss) / $0.014 cache hit / $1.32 output per 1M, V4 Pro peak $1.32 input / $0.044 cache hit / $3.96 output per 1M; off-peak 50% off (01:00–04:00, 06:00–10:00 UTC); effective 2026-08-16 16:00 UTC
  • Qwen CodeCLI is free/open source; hosted free tier restricted to ~100 requests/day as of April 2026; Coding Plan Pro ~$50/month (~90K requests/month); BYOK available
  • Total$47/mo

Why this stack exists

As of mid-2026 the Chinese coding line (Kimi K3, GLM-5.x, Qwen3.x, DeepSeek V4) trades blows with the US frontier on the vote-based coding boards while costing 2–5x less per month. If your bottleneck is budget, not capability, this is the sane default.

The recipe

  1. Daily driver: Kimi Code (membership ~$19–39/mo). The terminal agent for feature work and refactors. K3 is the strongest open-weights coder on the boards — long-horizon sessions hold up, and the CLI itself is free/MIT if you ever want to self-host the weights.
  2. Always-on second seat: GLM Coding Plan (Lite ~$18/mo). Use it for autocomplete-class work, small fixes and “review this diff” passes where burning Kimi allowance is waste. Lite is cheap enough to leave running all day.
  3. Bulk and batch: DeepSeek API (~$10/mo pay-as-you-go). Mass renames, test generation, log triage — anything scriptable goes through the API, scheduled into the 50%-off off-peak windows (01:00–04:00 and 06:00–10:00 UTC). Cache-hit pricing makes repetitive prompts nearly free.
  4. Free backup: Qwen Code (free tier). Open-source CLI, ~100 requests/day on the free tier, BYOK when you need more. Keeps you moving on the days a subscription lapses or a provider wobbles.
  5. One spec file, all agents. Keep AGENTS.md in the repo root — every one of these CLIs reads it. Write conventions once, get consistent behavior across all four tools.
  6. Route by stakes. Throwaway scripts and bulk chores → DeepSeek/GLM. Features you’ll maintain → Kimi. The 2 a.m. production incident → whatever is strongest that day, cost be damned.

Working with these tools, honestly

  • Prompts port cleanly. All four CLIs speak the same “spec + repo + small steps” dialect. No per-vendor prompt magic needed.
  • Context is huge, use it. K3 and Qwen3 both take 1M tokens — paste whole modules instead of hunting for the perfect snippet.
  • Payments are the real friction. Some plans need regional workarounds or annual billing for the good price; factor the setup hour into your decision.
  • Read the data terms. If your code can’t leave your jurisdiction, run the open weights (K3, GLM-5.2, Qwen3 are all downloadable) or skip this stack entirely.

What you pay

  • Kimi Code — hosted membership, ~$19/mo. The low end of the regional membership range; enough for daily feature work and refactors. The CLI is free if you bring your own keys.
  • GLM Coding Plan — Lite, ~$18/mo. Flat-rate autocomplete and small fixes; cheaper than burning Kimi allowance on low-stakes keystrokes.
  • DeepSeek API — pay-as-you-go, ~$10/mo. Bulk renames, test generation and log triage scheduled into off-peak windows; cache-hit pricing keeps repetitive prompts nearly free.
  • Qwen Code — free tier, $0/mo. ~100 requests/day on the hosted free tier or BYOK for more; the CLI itself is open source.

When this breaks down

Regulated codebases, strict data-residency requirements, or teams that need US-based support contracts — that’s the American stack’s home turf. And if you hit a task the Chinese line genuinely can’t finish (it happens, mostly on obscure frameworks), keep one Claude Code seat as the tiebreaker — see the Mixed stack.

Sources

  1. 1codeagentswarm.com
  2. 2aipricing.guru
  3. 3api-docs.deepseek.com
  4. 4github.com

Edit this page on GitHub