The open-source AI landscape in August 2026 looks nothing like it did in March. Qwen 3.8 Max launched with 2.4T params and 16-day autonomous coding. Kimi K3 is the largest open-weight model ever. DeepSeek V4 Pro holds SWE-bench records. And GPT-5.6 Luna’s price drop changed the economics entirely.

Here’s the current state of open-source AI models for August 2026.

The Contenders

October 20 additions: open weights without automatic ranking promotion

Muse Spark Open is now downloadable: 816B total/42B active parameters, multimodal inputs and a 1,048,576-token context. Its Muse Model License is not MIT/Apache; review attribution, use restrictions and the separate agreement requirement for qualifying high-revenue Model-as-a-Service businesses. This is a distinct self-hosting option, not a retroactive claim that the hosted Muse Spark releases were open source.

NVIDIA’s October 15 release adds Nemotron 3.5 Super VL and Bolt 74B-A9B downloadable weights. Bolt E4B was announced for October 23, not available at this check. Runtime support and vendor efficiency benchmarks make these candidates to evaluate, not independently demonstrated winners over the existing shortlist. See the Nemotron family guide and RTX Spark model-fit guide for deployment constraints.

	Qwen 3.8 Max	Kimi K3	DeepSeek V4 Pro	Qwen 3.7 Max	GLM-5.1
Company	Alibaba	Moonshot AI	DeepSeek	Alibaba	Zhipu AI
Total params	2.4T	2.8T	1.6T MoE	Not disclosed	754B MoE
Active params	95B	~200B est.	49B	Not disclosed	40B
Context window	1M	1M	1M	1M	1M
License	Open-weight (next week)	Open-weight	MIT	Proprietary (API)	MIT
API price (in/out)	~$3-5/$10-15 est.	$3/$15	$2.19/$8.76	$2.50/$7.50	$3/mo plan

Best overall: Qwen 3.8 Max

Qwen 3.8 Max is Alibaba’s newest flagship. 2.4T total params with 95B active, multimodal (text + vision), and 16-day autonomous coding capability.

Key strengths:

• Text Arena: #5 overall
• Vision Arena: #2 overall
• Frontend Code Arena: #4 overall
• 16-day autonomous coding (built oh-my-cli from scratch)
• Open weights coming next week

The pricing has not been published but is expected at $3-$5/$10-$15 per 1M tokens.

Best open-weight (available now): Kimi K3

Kimi K3 is the largest open-weight model ever released. At 2.8T parameters, it scores 88.3% on Terminal-Bench, second only to GPT-5.6 Sol among all models.

Key strengths:

• Terminal-Bench 2.1: 88.3% (#2 overall)
• Artificial Analysis Intelligence Index: #3 (near Opus 4.8)
• Always-on reasoning mode
• 1M context window
• Open weights (1.56 TB on HuggingFace)

The $3/$15 API pricing is mid-tier. Self-hosting requires serious hardware (multiple 80GB+ GPUs). For most developers, the API is the practical option.

Best for software engineering: DeepSeek V4 Pro

DeepSeek V4 Pro holds the SWE-bench Verified record at 80.6%. It’s the best open model for software engineering tasks.

Key strengths:

• SWE-Bench Verified: 80.6% (record)
• Codeforces: 3206 (#23 among all human competitors)
• LiveCodeBench: 93.5%
• MIT license (most permissive)
• $2.19/$8.76 API pricing

The MIT license allows unrestricted commercial use, fine-tuning, and redistribution. That matters for enterprises with strict licensing requirements.

Best value: Qwen 3.7 Max

Qwen 3.7 Max offers frontier-class performance at mid-tier pricing.

Key strengths:

• 1M context window
• Strong instruction following
• $2.50/$7.50 API pricing (cheapest of the frontier models)
• Expected open weights later

The main limitation: no open weights yet. If you need self-hosting, Kimi K3 or DeepSeek V4 Pro are the options.

Best for autonomous coding: GLM-5.1

GLM-5.1 was the first open-source model to top SWE-Bench Pro. Its 8-hour autonomous coding capability is unmatched.

Key strengths:

• SWE-Bench Pro: 58.4% (#1 at launch)
• 8-hour autonomous coding sessions
• MIT license
• $3/month coding plan

The MIT license and $3/month coding plan make it the most accessible frontier model.

Best for token efficiency: MiMo V2.5 Pro

MiMo V2.5 Pro achieves 57.2% on SWE-bench Pro with 40% fewer tokens than Opus 4.8.

Key strengths:

• 57.2% SWE-bench Pro
• 40% fewer tokens than Opus 4.8
• Apache 2.0 license
• $0.80/$3.20 API pricing

If you optimize for tokens consumed rather than raw accuracy, MiMo is the efficiency king.

Best budget: GPT-5.6 Luna

GPT-5.6 Luna deserves mention despite being proprietary. At $0.20/$1.20 with 84.3% Terminal-Bench, it’s the economic choice for most tasks.

Key strengths:

• Terminal-Bench 2.1: 84.3%
• $0.20/$1.20 API pricing (after July 30 price drop)
• 1M context window

Not open-source, but at 10x cheaper than open-weight alternatives, it changes the economics of AI development.

Best for self-hosting: Qwen 3.6-27B

Qwen 3.6-27B is the best option for local deployment. At 27B dense parameters, it runs on a single GPU (22GB VRAM) with Apache 2.0 license.

Key strengths:

• 27B dense (runs on single GPU)
• 77.2% SWE-bench Verified
• Apache 2.0 license
• Free to run

If you need to run locally without API dependencies, this is the model.

How to Choose

“I want the best open-source AI model” Kimi K3 (frontier-class, open-weight) or DeepSeek V4 Pro (MIT license, best SWE-bench).

“I’m on a budget” GPT-5.6 Luna at $0.20/$1.20 for API. Qwen 3.6-27B for free self-hosting.

“I need to self-host” Qwen 3.6-27B (27B, single GPU) or DeepSeek V4 Flash (13B active, MIT).

“I need MIT license” DeepSeek V4 Pro, DeepSeek V4 Flash, or GLM-5.1.

“I need autonomous coding” GLM-5.1 (8-hour sessions) or Kimi K3 (frontier intelligence).

“I want the cheapest API” GPT-5.6 Luna ($0.20/$1.20) or DeepSeek V4 Flash ($0.14/$0.28).

Related Articles

• 10 Best Free AI Coding Models 2026
• Best Open-Source Coding Model 2026
• Qwen 3.8 Max Complete Guide
• Kimi K3 Complete Guide
• DeepSeek V4 Pro Complete Guide
• GPT-5.6 Luna Price Drop
• AI API Pricing Compared 2026

FAQ

What’s the best open-source AI model in 2026?

Kimi K3 is the largest and most capable open-weight model, scoring 88.3% on Terminal-Bench 2.1. For software engineering tasks specifically, DeepSeek V4 Pro holds the SWE-bench Verified record at 80.6% with a more permissive MIT license.

Are open-source models catching up to proprietary ones?

Yes, the gap has nearly closed. Kimi K3 ranks #3 on the Artificial Analysis Intelligence Index, close behind Claude Opus 4.8. For most practical tasks, open-weight models now match or exceed proprietary alternatives, though the very top frontier models (GPT-5.6 Sol, Opus 4.8) still lead on the hardest reasoning benchmarks.

Can I self-host these models?

Kimi K3 and DeepSeek V4 Pro both offer open weights you can self-host, though Kimi K3’s 2.8T parameters require multiple 80GB+ GPUs. For realistic self-hosting on a single consumer GPU, Qwen 3.6-27B (22GB VRAM) is the practical choice.

Which model should I use on a budget?

GPT-5.6 Luna at $0.20/$1.20 per million tokens is the cheapest capable API option, even though it’s not open-source. For a free, fully open alternative, Qwen 3.6-27B runs locally at no ongoing cost beyond your own hardware.