Chinese open-weight models in 2026: a production guide

Start with one number: $0.42. That's roughly what DeepSeek charges per million output tokens for V3.2 — a model that trades blows with Western flagships on code, math, and general reasoning. GPT-5.1's output list price is about $10; Claude Sonnet 5's is $15. The open-weight models don't win everywhere. But a price gap north of 20x forces a question every engineering team should be asking in 2026: which parts of your traffic actually need a flagship?

This guide profiles the five model families that matter right now — DeepSeek, Qwen, Kimi, GLM, and MiniMax — the way a colleague who runs them in production would: what each is good at, what it costs, where it falls over, and how to use them without adding a mainland-China endpoint to your compliance review.

Five families, five personalities

First, a definition. These are open-weight models: the weights are published, and dozens of inference providers — many US- and EU-based — serve them. "Chinese model" describes where the lab sits, not where your tokens go. That distinction matters later in this post.

DeepSeek V3.2 and R1: the price setters

DeepSeek is the reason this conversation exists. V3.2 is the general-purpose workhorse: a sparse-attention MoE design that makes long context cheap, with strong code, math, and structured-output performance. The list price is the headline — roughly $0.28 in and $0.42 out per million tokens. At that price the economics of failure change: you can retry, generate three candidates and pick the best, or over-fetch context instead of engineering around it. The R1 line adds explicit long-form reasoning; it's slower and chattier, so use it for the hard planning step, not the whole pipeline. Quirks: both trend verbose, occasionally fight terse formatting instructions, and reasoning traces can balloon output-token counts if uncapped. All fixable with a firm system prompt.

Qwen3-Max and the coder line: the all-rounder

Qwen3-Max is the most flagship-like of the group. Instruction following is tight, multilingual quality is the best on this list, and it's the model to hand a messy, underspecified prompt. The Qwen3-Coder variants are tuned for repo-scale agentic coding — long tool-use trajectories, not single-file completion. At roughly $1.20 in / $6 out it's the premium option here, and still about 40% below GPT-5.1 on output. Qwen also buys you a family: dozens of open-weight sizes with shared behavior, so a prompt tuned on Max degrades gracefully if you later self-host something smaller.

Kimi K2: the agent specialist

Moonshot trained Kimi K2 hard on tool calling and it shows: schemas get respected, arguments arrive well-formed, and it stays coherent deep into multi-step trajectories where other models start improvising. Context is in the 256k class, which matters when an agent drags an entire ticket history along. It's also one of the better prose stylists here for a non-reasoning model. List price is around $0.60 in / $2.50 out. If your workload is "model calls twelve tools, then writes a summary a human will read," start with K2.

GLM-4.6: the coding value pick

GLM-4.6 from Zhipu became the default engine in many coding-agent harnesses because it does the unglamorous parts well: applies diffs cleanly, follows repository conventions, and produces front-end code that renders. Context is 200k; list price is roughly $0.60 in / $2.20 out — call it 7x cheaper than Sonnet 5 on output, and competitive on a large share of everyday coding tasks. It's weaker as a general writer and can over-eagerly refactor code you didn't ask about. For a coding budget, that trade is usually easy.

MiniMax M2: cheap steps for busy agents

MiniMax M2 is built for cheap, fast steps. It's a small-activation MoE, so per-call latency and cost stay low — roughly $0.30 in / $1.20 out — which matters when an agent makes forty calls to finish one task. Agents live and die by cost per completed task, not cost per token, and M2's is excellent. It interleaves short bursts of reasoning with tool calls rather than front-loading one giant think block. Skip it for long-form writing or nuanced chat; as the inner loop of a high-churn agent, it's a bargain.

What they cost

Output tokens dominate the bill for agents and anything generation-heavy, so stare at the output column. Public list prices per million tokens at the time of writing:

Because these are open weights, multiple hosts compete to serve the same model, and routed prices are frequently below list. That's the wholesale-pricing effect a gateway passes through — current numbers are on the pricing page.

The honest limitations

Anyone calling these models a free lunch hasn't run them in production. The gaps are real — just narrower and more specific than the discourse suggests.

• Enterprise compliance is the biggest tax. Some customers contractually restrict model origin regardless of where inference runs. Open weights help — you can serve them from US or EU hosts with no data flowing to the originating lab — but security and legal review still takes weeks, not days. Budget for the paperwork, not just the tokens.
• Latency and quality depend on the host. First-party endpoints can be slow from the US and EU at peak. Third-party hosts differ in quantization (FP8 vs more aggressive schemes), context handling, and tool-call fidelity. Two endpoints both labeled "GLM-4.6" are not always the same experience. Measure per host, not per model.
• English long-tail edge cases persist. Idiom, niche cultural references, and strict brand-voice work still tilt toward the Western flagships. This rarely matters for code, extraction, or RAG synthesis; it can matter for consumer-facing chat with personality requirements. Qwen3-Max is closest to parity.
• Ecosystem friction is small but nonzero. Tool-calling dialects differ, reasoning-token accounting varies, and documentation quality is uneven. An OpenAI-compatible gateway hides most of this, but keep an eval suite so surprises show up in CI, not in front of users.

Which family for which workload

Opinionated defaults, based on what these families were optimized for. Your evals outrank this table.

The pattern that holds in practice: keep a flagship on the 5–10% of traffic where it demonstrably earns its price, and route the rest to the model that clears your quality bar at the lowest cost. Blended, that's where the 5–20x savings come from — not from replacing everything at once.

Routing around the downsides

Most of the limitations above are host problems, not model problems — and host problems are routable. INFRO serves these families through vetted providers, most US- and EU-hosted, and tracks each provider's latency, throughput, and error rate. Requests go to the fastest reliable host at the best price; if a provider degrades, traffic fails over automatically. You can pin or exclude providers per request — including requiring non-China hosting — without changing application code. All behind one OpenAI-compatible API, one key, and one bill, across 120+ models.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.infro.io/v1",  # the only line that changes
    api_key="YOUR_INFRO_KEY",
)

resp = client.chat.completions.create(
    model="deepseek/deepseek-v3.2",  # or "zai/glm-4.6", "moonshot/kimi-k2"
    messages=[{"role": "user", "content": "Summarize this ticket thread..."}],
)

Because the interface is identical across models, an A/B between Sonnet 5 and GLM-4.6 on your real traffic is a one-string change plus an eval run. That's the whole migration. Usage analytics then show, per route, whether the cheaper model is holding quality — so the decision stays empirical instead of ideological.

Try it on one workload

You don't need a migration project. Pick one high-volume, low-risk route — classification, extraction, RAG synthesis — point it at DeepSeek V3.2 or GLM-4.6 through INFRO, and compare the evals and the invoice after a week. The quickstart takes about five minutes: swap the base URL, keep your existing OpenAI SDK code. If the numbers don't hold, switching back is one string. They usually hold.