Verdict

Best Open-Weight LLMs 2026: DeepSeek V4, GLM-5.3, Kimi K3 and Qwen3.8

Our verdict: DeepSeek V4-Pro-0813 is the safest overall pick, with a plain MIT license and the cheapest official API here. GLM-5.3 and Kimi K3 are the coding and agent picks, and Qwen3.8 27B is the one you can run at home.

Four Chinese labs now give away models that trade blows with the best closed AI. DeepSeek, Z.ai (formerly Zhipu), Moonshot AI and Alibaba all shipped new open-weight releases between June and August 2026, and their own model cards put the top three within a point of each other on Terminal-Bench 2.1.

Open-weight AI models running on local hardware
Weights you can download: every model here can run on hardware you control, if you have enough of it. Illustration: TechVerdict.
Compared DeepSeek V4-Pro GLM-5.3 Kimi K3 Qwen3.8 27B Qwen 3.5

How big these models are

Infographic: Kimi K3 2.8T total and 104B active, DeepSeek V4-Pro 1.65T total and 49B active, GLM-5.3 753B total with active not stated, Qwen3.8 27B dense with 27B active
Total vs. active: sizes from each lab's own Hugging Face model card (DeepSeek, Z.ai, Moonshot AI, Alibaba Qwen), checked Oct 23, 2026. Infographic: TechVerdict.

Total size decides how much memory you need. Active size decides speed, because a Mixture-of-Experts model only uses a slice of itself for each word. That's why a 2.8T model can answer quickly on a server, yet still won't fit on any desktop.

The 5 models at a glance

DeepSeek V4-Pro-0813
Best overall
1.65T 49B active
  • MIT license
  • 1M token context
  • API from $0.66/$1.98 per M tokens off-peak

The catch: text only, and server-sized.

GLM-5.3
Best for agentic coding
753B active not stated
  • 88.2 on Terminal-Bench 2.1
  • 66.9 on DeepSWE
  • MIT-style GLM-5.3 license

The catch: $10B+ API providers need a security review.

Kimi K3
Best for long agent sessions
2.8T 104B active
  • Text, image and video input
  • 1M token context
  • 88.3 on Terminal-Bench 2.1

The catch: custom license, and the biggest here.

Qwen3.8 27B
Best to run at home
27B dense
  • Apache 2.0
  • Reads images and video
  • 262K context, extensible to 1M

The catch: a step below the giants on hard coding.

Qwen 3.5
Best for multilingual
397B 17B active
  • Apache 2.0
  • 201 languages and dialects
  • Built-in vision encoder

The catch: older, and still big-workstation sized.

Specs and licenses

Specs checked October 23, 2026, on each lab's own Hugging Face model card. API prices are from DeepSeek's own pricing page.

Model (newest)Size (total / active)LicenseWhere it runsBest for
DeepSeek V4-Pro-08131.65T / 49BMITMulti-GPU server or DeepSeek API ($0.66/$1.98 per M tokens off-peak)Overall capability at low API cost
GLM-5.3753BMIT-style, with a security review for very large API providersMulti-GPU serverAgentic coding
Kimi K32.8T / 104BKimi K3 License (custom)Multi-GPU serverLong agent sessions, multimodal
Qwen3.8 27B27B denseApache 2.024GB GPU at 4-bit, or a 32GB+ MacRunning at home
Qwen 3.5397B / 17BApache 2.0Big workstation or serverMultilingual (201 languages and dialects)

What each one can do

FeatureV4-ProGLM-5.3Kimi K3Qwen3.8 27BQwen 3.5
Runs on a Mac✗✗✗✓Mac Ultra
Commercial license✓$10B ruleWith terms✓✓
Vision (image input)✗✗✓✓✓
Coding (Terminal-Bench 2.1)87.988.288.373.0–

– means the lab hasn't said. Scores are vendor-reported. Kimi K3's license asks API businesses over $20M a year to sign a separate deal. GLM-5.3-Flash, the smaller sibling, does take images.

DeepSeek V4-Pro: the overall leader

V4-Pro launched in April as a 1.6T Mixture-of-Experts model with 49B active, MIT-licensed, with a 1M token context. DeepSeek's card lists 80.6% on SWE-Bench Verified and 90.1% on GPQA Diamond for the Max variant. The August release, V4-Pro-0813, reports 87.9 on Terminal-Bench 2.1.

The API is the draw: $0.66 per million input tokens and $1.98 output off-peak, double that at peak. That's about 15x cheaper on input and 25x cheaper on output than Claude Fable 5.1 ($10/$50).

DeepSeek chart comparing V4-Pro-Max with Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro on seven benchmarks, plus charts showing lower compute and KV cache than DeepSeek V3.2
V4-Pro at launch: level with the closed leaders on SWE Verified, with far less compute at 1M tokens. Image: DeepSeek.

GLM-5.3: the coding specialist

GLM-5.3 keeps GLM-5.2's 753B base and gets all its gains from post-training. Z.ai's card reports 88.2 on Terminal-Bench 2.1 and 66.9 on DeepSWE, and Z.ai calls it the most capable open-weight model for coding.

Its license is MIT-style with one condition: API providers earning over $10 billion a year must pass a Z.ai security review. GLM-5.3-Flash (320B, 18B active) is plain MIT.

Z.ai bar charts comparing GLM-5.3 with GLM-5.2, Kimi K3, Fable 5 and GPT-5.6 Sol on six benchmarks
GLM-5.3 vs. the field: Z.ai's own chart, with GLM-5.3 ahead on AutomationBench and GDPval. Image: Z.ai.
Try before you buy hardware: run your own prompts through OpenRouter or Together AI first. Benchmarks here are vendor-reported, and the right model for your codebase may not be the one with the top score.

Kimi K3: the long-session agent

Kimi K3 is the giant: 2.8T total with 104B active, a 1M token context, and native text, image and video input. Moonshot's card reports 88.3 on Terminal-Bench 2.1, and Moonshot pairs it with its Kimi Code CLI for long engineering sessions.

It ships under a custom Kimi K3 License rather than MIT. If you want a smaller Moonshot model with a Modified MIT license, Kimi K2.6 (1T, 32B active) is still available.

A rider on a white horse in a procedural 3D open-world village built by Kimi K3
Built by K3: a browser-based 3D open-world game. Image: Moonshot AI.
A ray-traced black hole simulation with a control panel, recreated by Kimi K3
Vision in the loop: K3 checked its own screenshots to recreate Gargantua. Image: Moonshot AI.

Qwen3.8 27B and Qwen 3.5: the practical picks

Qwen3.8 27B is the model most people can actually run. It's dense, Apache 2.0, reads images and video, and handles 262K tokens natively (extensible to 1M). It fits a 24GB GPU at 4-bit or a 32GB+ Mac.

Qwen 3.5 (397B, 17B active, Apache 2.0) remains the multilingual pick, with 201 languages and dialects. Apache 2.0 is also the license enterprise legal teams know best.

What hardware you need

MemoryWhat runs
16GB MacStrong 9B-to-12B low-bit models such as Gemma 4 12B
24GB GPU or 32GB to 64GB MacQwen3.8 27B, plus older 70B-class quantizations at 64GB
128GB MacStep 3.5 Flash and 119B-to-122B models
192GB Mac or multi-GPUThe original DeepSeek V4-Flash (roughly 160GB mixed FP4/FP8 checkpoint)
Multi-GPU serverV4-Pro, GLM-5.3, Kimi K3, V4.1-Flash, Qwen3.8 2.4T

Full V4-Pro needs roughly 800GB of VRAM at full precision, about 10 H100 80GB GPUs. For every consumer tier from 16GB to 192GB, see our best local LLM by RAM guide.

Three ways to run them

PathCostGood forYou give up
Managed API (Together AI, Fireworks, DeepInfra, OpenRouter)Pay per tokenMost users, no setupPrivacy: the host sees your prompts
Rented GPUs (RunPod, Lambda, Vast.ai)About $2 to $4.30 per H100 hourYour own instance on demandSetup time and cold starts
Own hardware (2-4 GPU workstation or Mac Studio Ultra)$6,000 to $25,000 upfrontDaily use, full data controlFlexibility if architectures change

For serving, use Ollama or LM Studio on a Mac or gaming PC, and vLLM or SGLang on a server. The DeepSeek, GLM and Kimi cards all include vLLM or SGLang launch recipes.

What about V4-Flash, V4.1-Flash and Kimi K2.6?

DeepSeek V4-Flash (284B, 13B active) is the original home-lab DeepSeek, workable at 192GB. V4.1-Flash is a different beast at 763B, MIT, 90.6 on Terminal-Bench 2.1, and DeepSeek's cheap API tier at $0.15/$0.60 off-peak. Kimi K2.6 was the first open model to beat GPT-5.4 on SWE-Bench Pro, at 58.6.

Which one should you use?

You have a 24GB GPU or a 32GB+ Mac
Qwen3.8 27B
You have a 192GB Mac
DeepSeek V4-Flash
You want the best value API
DeepSeek V4-Pro-0813
You write code with agents
GLM-5.3
You need images, video and long sessions
Kimi K3
You work across many languages
Qwen 3.5

The verdict

DeepSeek V4-Pro-0813 is the best open-weight LLM for most teams in 2026: MIT-licensed, near the top on coding, and cheap on DeepSeek's own API. If you want one on your own machine, start with Qwen3.8 27B.

FAQ

Are Chinese open-source LLMs actually better than GPT-5.4?

On several benchmarks, yes. At release, GLM-5.1 scored 58.4 on SWE-Bench Pro versus GPT-5.4's 57.7 and Claude Opus 4.6's 57.3, and Kimi K2.6 scored 58.6, the first open-weight model to beat GPT-5.4 (xhigh) there. The newest releases push further: DeepSeek V4.1-Flash reports 90.6 on Terminal-Bench 2.1, Kimi K3 reports 88.3 and GLM-5.3 reports 88.2. These are vendor-reported scores, so treat them as a guide rather than a final ranking.

What hardware do I need to run GLM-5 or DeepSeek V4 locally?

It depends on size. The original DeepSeek V4-Flash (284B, 13B active) has an official mixed FP4/FP8 checkpoint of roughly 160GB before runtime and context, so 192GB of unified memory or a large multi-GPU setup is the practical starting point. The newer V4.1-Flash, DeepSeek V4-Pro, GLM-5.3 and Kimi K3 are much larger multi-GPU server models. Smaller dense models such as Qwen3.8 27B fit on a 24GB consumer GPU at 4-bit.

What's the best open-source LLM in 2026?

For overall capability, DeepSeek V4-Pro (now the 0813 release, MIT license, 1M context) is still the safest pick. For coding and agent work, GLM-5.3 and Kimi K3 are the newest contenders, both reporting about 88 on Terminal-Bench 2.1. For multilingual work, Qwen 3.5 (Apache 2.0, 201 languages and dialects). For a model you can run at home, Qwen3.8 27B (Apache 2.0). There's no single winner; the answer depends on your workload.

Can I actually run these models on a Mac?

Yes, but tier matters. A 16GB Apple Silicon Mac can run strong 9B-to-12B low-bit models. A 32GB to 64GB Mac runs Qwen3.8 27B comfortably and has room for older 70B-class quantizations. The original DeepSeek V4-Flash starts in roughly 192GB territory, while GLM-5.3, Kimi K3 and DeepSeek V4.1-Flash remain multi-GPU server models even at low precision.

How do I access these models if I can't run them locally?

Three paths. Managed APIs: DeepSeek's own API lists V4-Pro at $0.66 per million input tokens and $1.98 per million output tokens off-peak, and providers like Together AI, Fireworks, OpenRouter and DeepInfra host the major open models. Cloud GPU rental: RunPod and Lambda list H100s at roughly $2 to $4.30 an hour. Home cluster: a workstation with 2-4 GPUs or a Mac Studio Ultra gets you into large MoE territory. Managed APIs are the sanest starting point for most users.