Open-weight AI models: 88 you can run on your own hardware

Prices from official vendor price lists · October 2026
88 models with downloadable weights — self-host, fine-tune and stop paying per token. Output prices run from $0.059 to $15 per 1M tokens across the 82 models with a paid per-token price; 57 have a free tier.
Open weights
Services found: 88
#AI modelCapabilitiesAPI $/1M (in / out)Context windowPrice / moOutput $/1MOverviewVisit site
81

MiniMax M2-her

MiniMax
Text & chatReasoningTool calling
$0.3 / $1.265.5K tokensAPI only1.2View
82

MiniMax M2

MiniMax
Text & chatReasoningTool calling
$0.255 / $1.02204.8K tokensAPI only1.02View
83

Qwen3 Max Thinking

Qwen
Text & chatReasoningAdjustable thinking
$0.78 / $3.9262.1K tokensAPI only3.9View
84

Llama 4 Maverick

Meta
Text & chatVisionLong context 1M+
$0.2 / $0.6961.05M tokensAPI only0.696View
85

Llama 4 Scout

Meta
Text & chatVisionLong context 1M+
$0.1 / $0.310M tokensAPI only0.3View
86

Hy4 preview

Tencent
Text & chatReasoningAdjustable thinking
$0.84 / $2.521M tokensAPI only2.52View
87

Ling 3.0 Flash VL

InclusionAI
Text & chatVisionVideo
$0.02 / $0.059256K tokensAPI only0.059View
88

Ternary Bonsai 2 27B

PrismML
Text & chatVisionReasoning
262.1K tokensAPI onlyView

Browse by need

Reasoning models169 modelsCoding assistants17 modelsVision and images156 modelsLong context 1M+88 modelsFree tier available80 modelsOpen weights88 models

Ninety-four models in this rating publish downloadable weights, so you can run them on your own hardware, fine-tune on your own data and stop paying per token. The families to know: GLM, Qwen, DeepSeek, Kimi, Mistral Apache-licensed range, Gemma 4, Muse Glimmer 30B, Granite, EXAONE, Solar, Nemotron and Phi.

Hosted access still exists for most of them at $0.18 to $15 per million output tokens, so you can prototype against an API and move to self-hosting once volume justifies the GPUs. Read the licence before you plan: open weights covers everything from Apache 2.0 to community licences with usage thresholds and naming requirements.

You own the runtimeSelf-host, fine-tune, run air-gapped — no per-token bill at all.
Licence is not uniformApache 2.0 and restrictive community terms both sit under this label.
Hosted first, then ownPrototype on an API, move to GPUs once volume justifies it.

Models with this capability

The 88 models below all support this capability. Prices are input and output per 1M tokens from the vendor official price list; for dirhams multiply by 3.6725. Tap a card to open the model page.

DeepSeek V4.1 Flash
#1
DeepSeek
in $0.3out $1.21M tokens
GLM 5.3
#2
Z.ai
in $1.4out $4.41.31M tokens
Kimi K3
#3
Moonshot AI
in $3out $151.05M tokens
GLM 5.3 Flash
#4
Z.ai
in $0.075out $0.251.31M tokens
Qwen3.8 Max (0902)
#5
Alibaba (Qwen)
in $2out $61M tokens
Qwen3.8 2.4T A95B
#6
Alibaba (Qwen)
in $2out $61.05M tokens
DeepSeek V4 Pro 0813
#7
DeepSeek
in $1.0494out $3.14821.05M tokens
DeepSeek V4 Flash 0731
#8
DeepSeek
in $0.065out $0.181.31M tokens

How to choose

  • Read the licence for usage thresholds, naming requirements and field-of-use limits.
  • Compare total cost honestly: GPUs, ops and idle capacity against per-token rates.
  • Check whether the published weights match the hosted version you tested.
  • Plan for updates — self-hosted models do not improve on their own.

Common mistakes

Self-hosting too earlyAt prototype volumes the GPUs cost more than the API ever would.
Assuming open means unrestrictedSeveral popular licences restrict commercial use above a size threshold.

Who this suits

Regulated industriesData that cannot leave the perimeter under any terms.
High steady volumeWhere per-token billing has become the largest line item.
ResearchReproducibility and the freedom to inspect and modify.

Related filters

Proprietary API151 closed-weight, API-only models
Free tier80 models with a free API allowance
Low latency73 small, fast models for volume
Coding17 models built for writing code
All 239 modelsThe full rating with filters and prices

Frequently asked questions

Which families publish weights?

GLM, Qwen, DeepSeek, Kimi, Mistral Apache-licensed models, Gemma 4, Muse Glimmer 30B, Granite, EXAONE, Solar, Nemotron and Phi.

Is open weights the same as open source?

No. Licences range from Apache 2.0 to community terms with usage thresholds and naming requirements. Read before you build.

When does self-hosting pay off?

At steady high volume. Below that, hosted access from $0.18 per 1M output tokens is cheaper than the GPUs.

Do open models match the frontier?

Not yet on the hardest tasks. On routine work the gap is small and shrinking.

Can I fine-tune them?

Yes — that is the main practical advantage over closed models.

Frequently searched

By capabilityReasoningCodingVisionLong context 1M+Free tierOpen weightsLow latencyRealtime voiceImage generation
By vendorAnthropic (13)OpenAI (30)Google DeepMind (17)xAI (8)Meta (6)DeepSeek (9)Mistral AI (11)Alibaba (Qwen) (34)Z.ai (12)Moonshot AI (5)
Popular modelsClaude Fable 5.1GPT-6 AstraClaude Opus 5Claude Fable 5GPT-5.6 SolGLM 5.3Grok 4.6Kimi K3GPT-5.6 Terra
By deliveryBatch APIPrompt cachingPriority tierFrontier tierProprietary APITool callingStructured JSONWeb searchMultilingual

Ratings 1

Avg. score 5

Rate this page

Help us improve