Kimi K3 vs GLM-5.3: Open-Weight 1M-Context Flagships

All figures below are from the China AI Hub database, last verified 2026-09-20. Where a field is not publicly disclosed, we say so rather than estimating.

At a glance

Dimension	Kimi K3	GLM-5.3
Status (database)	active	active
Context window	1,048,576 tokens	1,048,576 tokens
Maximum output	1,048,576 tokens	131,072 tokens
Input price (per 1M)	$3.00	$1.40
Output price (per 1M)	$15.00	$4.40
Open weights	Yes (Kimi K3 License)	Yes (Apache-2.0)
API available	Yes	Yes
Reasoning	Yes	Yes
Coding	Yes	Yes
Vision / Video	Yes / Yes	No / No
Tool calling	Yes	Not listed
Structured output	Yes	Not listed
Agent capability	Yes	Not listed
Benchmark records in DB	6	4

Pricing

Kimi K3 lists $3.00 input / $15.00 output per 1M tokens; GLM-5.3 lists $1.40 input / $4.40 output — a bit over half the input price and under a third of the output price. Both are official pay-as-you-go API prices. The output-price gap (roughly 3.4x) is the more material figure, because reasoning workloads multiply output tokens and K3’s 1M-token output ceiling makes a single maximal response a material expense in absolute terms.

Context and output

Both list a 1,048,576-token context window. The output ceilings differ sharply: Kimi K3 lists the full 1,048,576-token maximum output — the only model in the database documented with 1M in and 1M out — while GLM-5.3 lists 131,072. For long-form generation this is the decisive listed difference between the two.

Coding

Both list coding. Kimi K3 lists vendor-reported Terminal-Bench 2.1 (88.3) and DeepSWE (67.5); GLM-5.3 lists Terminal-Bench 3.0 (28.3) and DeepSWE v1.1 (66.9). The benchmark versions differ (Terminal-Bench 2.1 vs 3.0), so the two coding numbers are not directly comparable as a head-to-head. Kimi K3 additionally lists tool calling and structured output where GLM-5.3 lists neither.

Vision

This is a clear listed split: Kimi K3 lists native vision and video input; GLM-5.3 is text-only input. For image or video understanding workflows, only K3 lists the capability. The caveat on K3 is that the exact input-modality boundary is itself inconsistently documented across Moonshot’s sources (architecture table says Text+Image; README, blog and API guide also list video).

Agent capabilities

Kimi K3 lists agent capability, tool calling and structured output; GLM-5.3 lists none of these three. For agent orchestration that depends on a documented tool/structured-output surface, the database records a difference in listed capability.

License and openness

Both are open weight, but the licenses differ materially. GLM-5.3 is Apache-2.0 (per GitHub repo metadata, with a documented caveat that the README has no separate weights-license section — verify per-model Hugging Face cards before reuse). Kimi K3 uses the Kimi K3 License: MIT-style permissive, but Model-as-a-Service operators above $20M aggregate revenue must sign a separate agreement, and products above 100M MAU or $20M monthly revenue must display “Kimi K3” in the UI. Apache-2.0 carries an explicit patent grant that the Kimi K3 License does not state; the Kimi K3 License carries revenue-triggered obligations that Apache-2.0 does not.

Deployment

Both are self-hostable open weight. Both are API-available. GLM-5.3 ships open weights in FP8 and BF16 and lists regions international and china; Kimi K3 lists a Moonshot AI Platform cloud provider. A self-hosting decision between them turns on the license terms and the serving footprint (K3 is a 2.8T-total MoE; GLM-5.3 is 744B-total) more than on availability.

Benchmarks

The database holds 6 benchmark records for Kimi K3 (GPQA Diamond, HLE-Full, DeepSWE, Terminal-Bench 2.1, MMMU-Pro, Video-MME) and 4 for GLM-5.3 (Terminal-Bench 3.0, DeepSWE v1.1, Agents’ Last Exam CLI, CyberGym). The overlap is minimal. All are vendor-reported. Counts are not scores.

Benchmark comparability is limited: the two models’ scores come from different test configurations — different benchmark versions (Terminal-Bench 2.1 vs 3.0), evaluation harnesses and tool setups — and several benchmarks cover only one of the two. Do not rank on headline scores alone.

Why each difference matters

China AI Hub analysis indicates the following per-dimension implications, drawn from the listed facts above.

• Pricing: GLM-5.3 has an advantage at just over half of K3’s input price and under a third of its output price; the output-token gap compounds on reasoning workloads.
• Context: identical 1M-token input windows, but Kimi K3 has an advantage in long-form generation because it lists the full-window 1M-token maximum output versus GLM-5.3’s 131,072.
• Coding: both list coding, so neither has a listed capability advantage; the version mismatch (Terminal-Bench 2.1 vs 3.0) makes the recorded scores non-comparable.
• Vision / media: Kimi K3 is more relevant when image or video input is required — it lists vision and video where GLM-5.3 lists neither (text-only).
• Agent: Kimi K3 is more relevant when a documented tool-calling/structured-output/agent surface is core — it lists these where GLM-5.3 lists none.
• License: both open weight; the choice turns on Apache-2.0’s patent grant versus the Kimi K3 License’s revenue-triggered obligations — verify which you can operate under.
• Deployment: both self-hostable and API-available; GLM-5.3 lists a smaller serving footprint (744B total vs 2.8T), which may matter for self-hosting hardware.
• Benchmark: K3 records 6 and GLM-5.3 records 4 vendor-reported rows with minimal overlap and version mismatch; not directly comparable.

Trade-off summary

• Output length: Kimi K3 lists a 1M-token maximum output, the highest in our database; GLM-5.3 lists 131,072.
• Price: GLM-5.3 lists substantially lower input and output prices.
• Media and agent: Kimi K3 lists vision, video, tool calling, structured output and agent capability; GLM-5.3 lists none of these (text-only, reasoning always on).
• License: Apache-2.0 (GLM-5.3) versus Kimi K3 License with revenue-triggered obligations.

Choose by workload: long-form generation, multimodal and agent-heavy pipelines favor Kimi K3’s listed capabilities; text-only agentic/coding work and budget favor GLM-5.3’s price and simpler Apache-2.0 license. Verify current prices on the official pages before committing.

Decision context

China AI Hub analysis indicates the following decision-context implications, drawn from the listed facts above.

For API developers. GLM-5.3 lists lower input and output prices, and the output gap compounds on reasoning workloads. Kimi K3 lists tool calling, structured output and vision/video, which GLM-5.3 does not; K3 also lists a 1M-token maximum output for long generation. Latency is not publicly documented for either model on this page.

For self-hosting. Both are open weight and self-hostable. The license differs: Apache-2.0 (GLM-5.3, with patent grant, but verify per-model HF cards) versus the Kimi K3 License (revenue-triggered obligations above $20M / 100M MAU). Hardware requirements and quantization are Not publicly documented on this page; GLM-5.3 lists a smaller total-parameter footprint (744B vs 2.8T).

For coding agents. Both list coding. Kimi K3 lists tool calling and structured output plus a 1M output ceiling, which is more relevant for tool-driven, long-output agent loops; GLM-5.3 lists none of the tool/structured surface. The coding benchmark versions differ, so SWE/terminal scores are not directly comparable on this page.

For enterprise. Region, SLA, data-residency and compliance terms are not publicly documented — confirm with the vendor. Deployment is comparable (both self-hostable), but the license obligations differ: the Kimi K3 License has revenue thresholds, while GLM-5.3’s Apache-2.0 has a patent grant with a per-card verification caveat.

What is uncertain

• Benchmark comparability is limited — different versions, harnesses and minimal overlap.
• Kimi K3’s input-modality boundary is inconsistent across Moonshot’s official sources.
• GLM-5.3 lists no tool calling or structured output — an absence of documentation, not a verified lack of capability.
• No third-party benchmark evidence is recorded for either model.

Sources

• Moonshot AI — Kimi K3 (GitHub)
• Kimi platform — Models documentation
• Z.ai — GLM-5.3 documentation
• Z.ai — Pricing overview
• Z.ai — GLM-5 (GitHub)

Labels used above: Official fact (prices, context windows, capabilities and license terms from primary provider sources), Vendor-reported claim (benchmark scores), and China AI Hub analysis (the “why each difference matters” reasoning, introduced as analysis). No third-party benchmark evidence is currently recorded for these models.