The Best Open-Source LLMs in 2026
Seven open-weight model families compared on license, context window and the per-million-token price where a vendor lists one, checked on 10 October 2026.
Short answer: DeepSeek V4-Flash has the lowest paid API price of the seven, at $0.15 per million input tokens off-peak, and its weights are MIT-licensed. For MIT terms with a 1M-token context and a higher price, pick GLM-5.2 at $1.40 in and $4.40 out. For a model you run on your own servers under Apache 2.0, look at Qwen3.8-27B.
Three of the seven state a 1M-token context: DeepSeek V4, GLM-5.2 and Kimi K3. Output tokens cost three to six times the input rate wherever a vendor publishes both, so the output price matters more than the headline input rate. Read the license before you ship: MIT and Apache 2.0 are simple, and Kimi K3 uses its own license.
Key facts7 tools compared: DeepSeek V4, GLM-5.2 (Z.ai), Kimi K3 (Moonshot AI), Qwen3.8-27B (Alibaba)…
- 7 tools compared: DeepSeek V4, GLM-5.2 (Z.ai), Kimi K3 (Moonshot AI), Qwen3.8-27B (Alibaba), gpt-oss-120b (OpenAI), Gemma 4 (Google), Mistral Small 4 (Mistral AI)
- GLM-5.2 (Z.ai) (best for: Long, multi-step agent work where a 1M-token context and MIT terms matter more than price): MIT weights; API $1.40 in, $4.40 out per 1M tokens, $0.26 cached input
- Kimi K3 (Moonshot AI) (best for: Teams with the hardware to serve a 2.8-trillion-parameter model and a 1M-token context): Own Kimi K3 License; API $3 in, $15 out per 1M tokens, $0.30 cached input
- Qwen3.8-27B (Alibaba) (best for: Teams that want a multimodal model under Apache 2.0 that they can run on their own servers): Apache 2.0 weights; Alibaba Cloud API $0.50 in, $3 out per 1M tokens (International)
"Open-source" is loose here. Almost every model below is open-weight: you can download the parameters, but the training data stays private. The license decides whether you can build a paid product on the model, so this page lists it for each one.
We ranked the seven by what a buyer chooses on: price per token, license, context window, and how heavy the model is to run yourself.
Toolradar data: Toolradar tracks 42 AI model deployment tools as of October 2026, the software that serves a model on your own servers, and most of them have a free or open-source option.
Serving is the cost you carry when you self-host, so it belongs in the budget.
Toolradar profiles include Ollama, vLLM and OpenRouter, and the full list sits in the AI model deployment ranking.
How we ranked and checked: we read each vendor's own pricing page and each model card or license page on 10 October 2026.
Prices are USD per million tokens as the vendor lists them. We ran no benchmarks, so the order rests on price, license, context and size, not on scores. Where a vendor publishes no per-token price, the table says so.
Top Picks
Based on features, real-world fit, and value for money.
| Tool | Pricing | Best for |
|---|---|---|
| DeepSeek V4 | MIT weights; API $0.15 in, $0.60 out per 1M tokens (Flash, off-peak) | Teams that want the lowest paid API price among these models, with MIT weights to fall back on |
| GLM-5.2 (Z.ai) | MIT weights; API $1.40 in, $4.40 out per 1M tokens, $0.26 cached input | Long, multi-step agent work where a 1M-token context and MIT terms matter more than price |
| Kimi K3 (Moonshot AI) | Own Kimi K3 License; API $3 in, $15 out per 1M tokens, $0.30 cached input | Teams with the hardware to serve a 2.8-trillion-parameter model and a 1M-token context |
| Qwen3.8-27B (Alibaba) | Apache 2.0 weights; Alibaba Cloud API $0.50 in, $3 out per 1M tokens (International) | Teams that want a multimodal model under Apache 2.0 that they can run on their own servers |
| gpt-oss-120b (OpenAI) | Apache 2.0 weights; no OpenAI API list price, so hosts set the rate | Teams that want OpenAI-made open weights under Apache 2.0 and will host them themselves or through a provider |
| Gemma 4 (Google) | Apache 2.0 weights; free on the Gemini API free tier, no paid rate listed | Small teams that want a 31B multimodal model under Apache 2.0 with a 256K context |
| Mistral Small 4 (Mistral AI) | Apache 2.0 license; no per-token rate on Mistral's pricing page | Teams already on Mistral's API that want one model for instruct, reasoning and coding |
Pricing read from each vendor's own published pricing page, checked Oct 2026. Every vendor here publishes a price.
Best for: Teams that want the lowest paid API price among these models, with MIT weights to fall back on
PricingMIT weights; API $0.15 in, $0.60 out per 1M tokens (Flash, off-peak)
Best for: Long, multi-step agent work where a 1M-token context and MIT terms matter more than price
PricingMIT weights; API $1.40 in, $4.40 out per 1M tokens, $0.26 cached input
Best for: Teams with the hardware to serve a 2.8-trillion-parameter model and a 1M-token context
PricingOwn Kimi K3 License; API $3 in, $15 out per 1M tokens, $0.30 cached input
Best for: Teams that want a multimodal model under Apache 2.0 that they can run on their own servers
PricingApache 2.0 weights; Alibaba Cloud API $0.50 in, $3 out per 1M tokens (International)
Best for: Teams that want OpenAI-made open weights under Apache 2.0 and will host them themselves or through a provider
PricingApache 2.0 weights; no OpenAI API list price, so hosts set the rate
Best for: Small teams that want a 31B multimodal model under Apache 2.0 with a 256K context
PricingApache 2.0 weights; free on the Gemini API free tier, no paid rate listed
Best for: Teams already on Mistral's API that want one model for instruct, reasoning and coding
PricingApache 2.0 license; no per-token rate on Mistral's pricing page
What it is
An open-weight LLM is a large language model whose trained parameters are published for download, so you can run it on your own hardware, through a cloud host, or both. Running it yourself removes the per-token bill of a closed API and replaces it with servers and staff.
Calling it through a vendor's API keeps the setup simple and charges per token. Most of the models here offer both routes, and the cheapest route depends on how many tokens you move each month.
If you only need a closed chatbot for daily work, our best AI chatbots list covers the paid options, and our private chatbots guide covers the data questions that self-hosting answers.
Why it matters
The price gap is the reason to read this list. Among the models with a paid API rate, DeepSeek V4-Flash is the cheapest on input and on output, and Kimi K3's output rate is 25 times the Flash off-peak output rate. A team sending 100 million input tokens a month sees that gap in its first invoice, not in a benchmark.
For the closed side of the bill, read our ChatGPT pricing, Claude pricing and Gemini pricing guides.
Key features to look for
Pricing
Every price here is a list price on the vendor's own page, checked on 10 October 2026. DeepSeek publishes rates for two V4 models and doubles them in its weekday peak windows, 01:00 to 04:00 and 06:00 to 10:00 UTC.
Z.ai and Moonshot show a cached-input rate beside the regular input rate, and Alibaba's rate applies to its International endpoint, covered in our Qwen3.8-27B news piece.
Hosts such as OpenRouter let developers reach many models through one API, as our OpenRouter news piece explains.
The scenario is 100 million input and 20 million output tokens a month, with no cache hits and off-peak pricing.
On that basis DeepSeek V4-Flash comes to $27 a month, V4-Pro to $105.60, Qwen3.8-27B to $110, GLM-5.2 to $228 and Kimi K3 to $600. Flash in a peak window doubles to $54.
Caching lowers these totals, which is why the cached-input column matters.
Sources: DeepSeek API pricing, Z.ai pricing, Kimi API pricing, Alibaba Cloud Model Studio pricing, Gemini API pricing, OpenAI API pricing and Mistral pricing.
| Plan | Price | Best for |
|---|---|---|
| DeepSeek V4-Flash, off-peak | $0.15 in / $0.60 out per 1M | Cache hit $0.003 per 1M input; 1M context; MIT weights |
| DeepSeek V4-Flash, peak | $0.30 in / $1.20 out per 1M | Weekday peak windows, 01:00 to 04:00 and 06:00 to 10:00 UTC |
| DeepSeek V4-Pro, off-peak | $0.66 in / $1.98 out per 1M | Larger model with no image input; cache hit $0.022 per 1M input |
| DeepSeek V4-Pro, peak | $1.32 in / $3.96 out per 1M | The same model in the weekday peak windows |
| GLM-5.2 (Z.ai) | $1.40 in / $4.40 out per 1M | Cached input at $0.26 per 1M; MIT weights; 1M context |
| Kimi K3 (Moonshot AI) | $3 in / $15 out per 1M | Cached input at $0.30 per 1M; 2.8T parameters; vision; 1M context |
| Qwen3.8-27B, Alibaba Cloud International | $0.50 in / $3 out per 1M | Thinking and non-thinking modes at one rate; a pricing tier for inputs up to 1M tokens |
| gpt-oss-120b (OpenAI) | No OpenAI list price | Hosts set the rate; weights are Apache 2.0 |
| Gemma 4 on the Gemini API | Free of charge | Free tier only; the paid tier is listed as not available |
| Mistral Small 4 | No per-token list price | Apache 2.0 weights; commercial terms may need a Mistral license |
The bottom line
For most buyers, DeepSeek V4-Flash is the first call: the lowest paid rate on this list, MIT weights you can self-host, and a 1M-token context.
Move up to GLM-5.2 when MIT terms and a long context matter more than price, and to Kimi K3 only if you can serve a 2.8-trillion-parameter model and accept its custom license.
For self-hosting under Apache 2.0, Qwen3.8-27B and Gemma 4 are the smaller choices.
If you want a closed model as the benchmark, see our ChatGPT review and Claude review.
Check prices again before you budget, because peak windows, cache rates and license terms change. Cite this: AIpresso, "Best Open-Source LLMs in 2026", October 2026, https://aipresso.com/reviews/best-open-source-llms.
Frequently asked questions
Some offers on this page may be paid placements or contain affiliate links.
Get the AIpresso brief
Free daily newsletter, read in 5 minutes.
Subscribe free