Best Chinese LLM for Coding (2026)

For a single coding turn, DeepSeek V4-Flash on its official endpoint is the cheapest credible option at roughly $0.0000374 per task, while qwen3-coder routed via OpenRouter ran fastest at 1.50 seconds for a near-identical $0.0000656. The real choice is not which model is cheapest in the abstract but which trades cost against latency in the way your coding workflow actually needs.

We answer that with one set of first-hand runs: a short coding prompt sent to five Chinese model families, billed and timed the same day. DeepSeek was called on its official api.deepseek.com endpoint. Qwen, GLM, Kimi, and MiniMax were measured via OpenRouter, so their per-token rates reflect that routing and carry a native-reverify caveat we flag below.

Coding cost and latency by model (verified 2026-06)

Here is the question most buyers actually have: for one realistic coding turn, what does it cost and how long does it take? We pulled the coding-labeled run for each family from our evidence pack and computed cost per task from the billed figures. DeepSeek's per-task cost is derived from its official $0.14 input / $0.28 output rate against the tokens we observed, since the official endpoint does not return a billed-cost field the way the routed calls do.

Model	Routing	Tokens (in/out)	Cost per task (USD)	Latency
DeepSeek V4-Flash	official api.deepseek.com	27 / 120	~$0.0000374	3.22s
qwen3-coder	OpenRouter	36 / 58	$0.0000656	1.50s
Kimi K2	OpenRouter	35 / 41	$0.0001143	4.55s
MiniMax-M2	OpenRouter	38 / 200	$0.0002514	2.95s
GLM-4.6	OpenRouter	30 / 165	$0.0002972	3.78s

Two caveats matter before you read too much into the cost column. First, output length drives most of the spread: MiniMax-M2 and GLM-4.6 simply generated longer answers (200 and 165 tokens) than qwen3-coder or Kimi K2 did, so part of their higher cost is verbosity, not a higher rate. Second, the DeepSeek figure is the only one billed at a native rate; the rest are OpenRouter pass-through and may differ on a direct vendor key.

What we measured first-hand (OpenRouter-routed, disclosed)

The numbers above are not scraped from a rate card. We called each endpoint with a coding prompt and recorded the billed cost and wall-clock latency the provider returned. For the four non-DeepSeek families, every call went through OpenRouter, and we report exactly what it billed.

According to our coding-run evidence pack, qwen3-coder returned 36 input and 58 output tokens for a billed $0.0000656 in 1.50 seconds, the fastest coding turn in the set. Kimi K2 came back in 4.55 seconds at $0.0001143 for a tight 35/41 token exchange, the cheapest of the routed pair after qwen. MiniMax-M2 took 2.95 seconds and billed $0.0002514 across 38/200 tokens, while GLM-4.6 ran 3.78 seconds at $0.0002972 for 30/165. The DeepSeek anchor call, on its official endpoint, returned 27/120 tokens in 3.22 seconds.

The pattern that surprised us was speed. The model people reach for on price, DeepSeek V4-Flash, was not the fastest here; qwen3-coder and MiniMax-M2 both beat it on this coding turn, with qwen more than twice as quick. Coding agents that fire many short calls in a loop feel that gap, so latency deserves a seat at the table next to cost.

DeepSeek V4-Flash is the cheapest credible coding tier

A coding-tier model is a general LLM whose price and speed make it economical to call repeatedly inside a code-editing loop rather than once for a long generation. By that definition, DeepSeek V4-Flash is the value anchor for the whole set.

According to DeepSeek API Docs, V4-Flash bills input at $0.14 and output at $0.28 per million tokens on the official endpoint. Against the 27 input and 120 output tokens we observed, that works out to about $0.0000374 for the turn, the lowest in our table even though its answer was longer than qwen's or Kimi's. The practical upshot is that on a per-token basis DeepSeek is hard to beat, and its native cache discount, covered in the DeepSeek API pricing hub, pushes a stable-prefix coding workload lower still.

The catch is the one the latency column already told you. DeepSeek wins on rate, not on response time, so if your bottleneck is how fast an agent iterates rather than the monthly invoice, the cheapest model is not automatically the right one.

qwen3-coder when speed is the bottleneck

When the workflow is an interactive coding agent, the metric that hurts is latency, and qwen3-coder is the standout. According to Alibaba Cloud, the Qwen3-Coder line is built specifically for agentic coding and tool use, which is consistent with what we saw: a 1.50-second turn, less than half the DeepSeek anchor's 3.22 seconds, for $0.0000656.

That cost sits just above DeepSeek and well below the GLM and MiniMax turns, so qwen3-coder is the model to choose when a developer is waiting on each completion and a few cents per thousand calls is an acceptable premium for halving the wait. It is the clearest both-fast-and-cheap pick in the routed group.

GLM-4.6 and Kimi K2 when output quality leads

GLM-4.6 and Kimi K2 are the higher-cost turns in our coding set, but cost per task is not the only axis. Both are positioned as strong reasoning-and-coding models, and their place here is for the workflow where a single high-quality completion is worth more than shaving fractions of a cent.

Kimi K2's 4.55-second, $0.0001143 turn was the slowest measured among the four routed families, so it suits batch or background coding tasks rather than tight interactive loops. GLM-4.6 landed mid-pack on latency at 3.78 seconds but highest on cost in this set, largely because it wrote the longest answer of the cheaper tier. According to MiniMax Platform, its M2 model targets coding and agent workloads, which fits the 2.95-second, $0.0002514 turn we saw and places it between the fast-cheap and quality-led groups. Neither GLM nor Kimi is the budget pick, and neither pretends to be.

Verdict by budget and workflow

The honest answer is that there is no single best Chinese coding model, only a best one for your constraint. Choose DeepSeek V4-Flash when total spend is the constraint and you can absorb a roughly 3-second turn. Choose qwen3-coder when an interactive agent makes latency the constraint and a small cost premium is fine. Reach for GLM-4.6 or Kimi K2 when a single strong completion matters more than per-call cost or speed, accepting the higher bill and, for Kimi, the longer wait. For the routed four, validate the rate on a native vendor key before you commit, because OpenRouter pricing is what we measured, not the vendor's direct rate.

FAQ

Which Chinese LLM is cheapest for coding? On our runs, DeepSeek V4-Flash was cheapest at roughly $0.0000374 per coding turn using its official $0.14/$0.28 rate, just under qwen3-coder's $0.0000656 measured via OpenRouter. DeepSeek wins on rate; qwen wins on speed.

Which Chinese coding model is fastest? qwen3-coder, measured via OpenRouter, returned its coding turn in 1.50 seconds, the fastest in our set and more than twice as quick as the DeepSeek V4-Flash anchor call at 3.22 seconds. MiniMax-M2 was next at 2.95 seconds.

Why is GLM-4.6 more expensive per task than Kimi K2 here? Mostly output length. GLM-4.6 generated 165 output tokens to Kimi K2's 41 on these coding prompts, so its $0.0002972 turn reflects a longer answer, not necessarily a higher underlying rate. Verbosity drives much of the cost spread.

Are these coding costs from official vendor endpoints? Only DeepSeek. Its numbers come from official api.deepseek.com. Qwen, GLM, Kimi, and MiniMax were measured via OpenRouter, so their per-task costs are routed figures and need native-key reverification before you treat them as the vendor's direct price.

Does cheapest per token mean cheapest for my coding workload? Not always. Output length and how many calls your agent fires both move the real bill, and a slower model that makes a developer wait has a cost the rate card never shows. Match the model to your binding constraint, cost or latency.

This is part of the best Chinese LLM API hub, which compares pricing across DeepSeek, Qwen, GLM, Kimi, and MiniMax. For one family's full rate card, see the DeepSeek API pricing hub.

Author: Kevin Fan, Customer Success Manager at China LLM Directory, specializing in Chinese LLM ecosystem pricing and model benchmarking. Last verified: 2026-06-26.