Kimi K3 is Moonshot AI's 2.8 trillion-parameter frontier model, announced July 16, 2026 alongside the World AI Conference (WAIC) 2026 in Shanghai. Moonshot committed to releasing open weights by July 27, 2026. K3 succeeds the widely-used Kimi K2.7 Code line and represents the largest publicly-committed open-weight release of 2026. Architecture details point to a sparse Mixture-of-Experts design with active parameters far below the 2.8T total.

How does Kimi K3 compare to DeepSeek V4 and GLM-5.2?

As of July 17, 2026, Kimi K3 is announced but weights are not yet public (target: July 27). DeepSeek V4 is fully open-weight and widely deployed on Huawei Ascend and NVIDIA H200 setups; V4 Pro is the current DeepSeek flagship for coding. GLM-5.2 (Zhipu) is the strongest general-purpose Chinese open-weight model with the tightest safety alignment. All three are Chinese labs, all three are competitive with GPT-5.5 and Claude Sonnet 5 on many benchmarks. Kimi K3's 2.8T total parameter size makes it the biggest — but bigger isn't automatically better.

Which open-weight Chinese model should I run in production?

DeepSeek V4 Pro today: proven at scale, mature vLLM/SGLang support, best coding benchmarks in the open-weight class. Kimi K3 once weights drop (July 27): worth evaluating for long-context and agent workloads if the sparse MoE routing works well on your hardware. GLM-5.2 for general enterprise chat with lower memory footprint. MiniMax M3 (open-weight, 1M-token context, native multimodal) for large-document and vision-heavy workloads. Pick based on workload — bigger total-parameter count is not the deciding factor.

Can I actually run Kimi K3 locally?

Not on consumer hardware. A 2.8 trillion-parameter model — even sparse MoE — requires cluster-scale inference. Realistic deployment: 8-16 H200 or Ascend 950 GPUs for the full model, or quantised distillations for smaller setups. For genuinely local use (single workstation), stick with 7B-70B distilled variants that will follow the K3 release, or use vLLM/SGLang on rented cloud clusters. Moonshot has previously released smaller distilled versions of the Kimi family; expect the same pattern here.

Quick Answer

Kimi K3 vs DeepSeek V4 vs GLM-5: Open Weights (Jul 2026)

Published: July 17, 2026

Kimi K3 vs DeepSeek V4 vs GLM-5.2: Chinese Open Weights (July 2026)

Moonshot AI announced Kimi K3 on July 16, 2026 — a 2.8 trillion-parameter frontier model with open weights committed by July 27, 2026, timed to the opening of the World AI Conference (WAIC) 2026 in Shanghai. It joins DeepSeek V4 Pro and GLM-5.2 as the three most-watched Chinese open-weight lines. Together with MiniMax M3, this is the strongest open-weight quarter since DeepSeek V3 in late 2024.

Here is what each is, how they compare, and which to pick for real deployments.

Last verified: July 17, 2026

The Chinese Open-Weight Line-up (July 2026)

Model	Lab	Total Params	Weights	Best At
Kimi K3	Moonshot AI	2.8T (sparse MoE, active far lower)	Announced July 16; open weights by July 27	TBD; likely long-context + agent
Kimi K2.7 Code	Moonshot AI	Undisclosed	Open (current)	Coding, currently widely used
DeepSeek V4 Pro	DeepSeek	~700B MoE (est.)	Fully open	Coding, math, cheap production inference
DeepSeek V4	DeepSeek	~400B MoE (est.)	Fully open	General reasoning, RAG
GLM-5.2	Zhipu AI	~200B dense + long-context variant	Fully open	Enterprise chat, tight alignment
MiniMax M3	MiniMax	Not disclosed; sparse MSA	Fully open	1M-token context, native multimodal
openPangu 2.0	Huawei	Optimised for Ascend	Partially open	Ascend chip stack

Kimi K3: What Is Announced

Moonshot AI’s July 16 announcement establishes:

Total parameter count: 2.8 trillion.
Architecture: Mixture-of-Experts (details expected with the weights release).
Open-weights commitment: by July 27, 2026 — 11 days after announcement.
Positioning: frontier tier; direct comparison to GPT-5.6, Claude Sonnet 5, Gemini 3.5 Pro on public benchmarks.
Predecessor: Kimi K2.7 Code has been the workhorse Chinese open-weight coding model since Q1 2026.

What is NOT announced yet (as of July 17): active parameters per token, exact context window, license terms for the weights, benchmark numbers with methodology, and specific quantisation-friendly variants.

DeepSeek V4 Pro: The Current King

DeepSeek’s V4 line is the most-deployed Chinese open-weight family in production:

V4 base and V4 Pro shipped earlier in 2026 with fully open weights and permissive licensing.
Mature tooling: vLLM, SGLang, TensorRT-LLM, MLC-LLM all have optimised paths.
Runs on both stacks: Huawei Ascend clusters and NVIDIA H200 (multi-node setups).
Best coding benchmarks in open-weight class, competitive with Claude Sonnet 5 on many tasks and cheaper to run than GPT-5.5 API.
Ecosystem: RAG frameworks, fine-tuning kits, quantised distillations widely available.

If you need to ship production inference today, DeepSeek V4 Pro is the safe pick.

GLM-5.2: The Aligned Enterprise Option

Zhipu AI’s GLM line is the corporate-safe choice:

Tightest safety alignment among Chinese open-weight models — heavy RLHF and content-moderation tuning.
Long-context variant with 128K-1M token context depending on release.
Dense architecture — easier to reason about VRAM requirements than MoE.
Enterprise-friendly license — deployed inside Chinese banks, telecoms, and government.
Weakness: slightly behind DeepSeek on raw benchmarks; the tradeoff for tighter alignment.

MiniMax M3: The Sleeper

MiniMax released M3 on June 1, 2026 and is showcasing it at WAIC 2026:

1 million token context window.
Sparse Attention (MSA) architecture — MiniMax’s proprietary efficient attention scheme.
Native multimodal — text, image, video inputs.
Claims to beat GPT-4.5 and Gemini 2.5 Pro on SWEbench Pro at significant cost advantage.
Fully open weights.

M3 is arguably the strongest launched-and-shipping open-weight coding model of Q3 2026 — with real benchmark data — while Kimi K3 is still an announcement.

Head-to-Head Decision Matrix

Use case	Best pick (July 17, 2026)
Production inference today	DeepSeek V4 Pro
Long context (>500K tokens)	MiniMax M3 (1M) or GLM-5.2 long-context
Multimodal (vision, video)	MiniMax M3
Coding benchmarks (open-weight)	DeepSeek V4 Pro or MiniMax M3
Enterprise-safe (aligned)	GLM-5.2
Ascend hardware	openPangu 2.0 or DeepSeek V4 on Ascend
Bleeding-edge experimentation	Kimi K3 once weights drop (July 27)
Cheapest cloud inference	DeepSeek V4 (via multiple providers)

The Bigger Story: Open-Weight Surge

WAIC 2026 timing is deliberate. The Chinese open-weight surge in 2026 has three drivers:

US chip export controls push Chinese labs to prove their models can run on Ascend and other non-NVIDIA hardware.
Diplomatic positioning. WAICO (formed July 16 with 29 signatory countries) and the “China Wisdom for the World” AI case collection want to position Chinese AI as the Global South alternative to closed US models.
Economics. Chinese labs monetise through cloud services, enterprise deployments, and government contracts — not through per-token API pricing at the frontier. Open weights are a distribution weapon.

Meanwhile in the closed-weight world: OpenAI’s GPT-5.6 Sol / Terra / Luna, Anthropic’s Claude Sonnet 5 / Opus 4.8, and (targeting July 17) Google’s Gemini 3.5 Pro. The gap between the frontier closed models and the best open-weight Chinese models is real but narrower than a year ago — a few points on major benchmarks, and closing.

Practical Deployment Guide

Single-workstation (24GB VRAM):

Quantised DeepSeek V4 distillations (7B-30B).
Quantised GLM-5.2 variants.
Do NOT try Kimi K3 or full DeepSeek V4 Pro locally.

Multi-GPU workstation (4-8× H200):

Full DeepSeek V4 or GLM-5.2 with vLLM.
MiniMax M3 on rented H200 nodes.
Quantised Kimi K3 (once distillations ship).

Production cluster (16+ GPUs):

DeepSeek V4 Pro with vLLM/SGLang for high throughput.
Kimi K3 once weights and reference deployments are validated.
Ascend-based deployments via Huawei’s stack for regulatory reasons.

Cloud (managed):

DeepSeek and MiniMax available via multiple hosted providers.
Kimi K3 will land on Moonshot’s own API and likely Groq/Together/DeepInfra/Fireworks within days of weights release.

What to Watch After July 27

Actual Kimi K3 benchmarks with methodology — SWE-bench Verified, LiveCodeBench, AIME, GPQA.
License terms — full open (Apache 2.0-style) vs research-only vs commercial-restricted.
Distillation cadence — how fast smaller variants ship.
Cursor / Aider / Continue integrations — routing share is the real popularity metric.
Enterprise deployment reference architectures — Moonshot’s own suggestions for large-scale inference.

Bottom Line

DeepSeek V4 Pro is the current open-weight production pick. MiniMax M3 is the current best long-context and multimodal open-weight. Kimi K3 is the biggest open-weight release of Q3 2026 by parameter count, but you need to wait for the July 27 weights drop and real benchmarks before switching workflows. GLM-5.2 remains the enterprise-safe option.

The Chinese open-weight scene is arguably the most interesting story of the summer — bigger for practical deployment than the frontier closed-model race, and largely under-covered outside the AI research community.

Sources

Moonshot AI: moonshot.cn news and Kimi K3 announcement pages
DeepSeek: deepseek.com
Zhipu AI GLM: zhipuai.cn and Hugging Face
MiniMax M3: minimax.io/blog/minimax-m3, minimax.io/models/text/m3
WAIC 2026 official portal: english.shanghai.gov.cn/en-WAIC2026