Unlimited use. Every open source LLM.
We recently launched LLMs, and you're one of a small group of teams we've invited to run every open source model on Runware with no credit cap and no token ceiling, free for 14 days. After that it rolls to pay-as-you-go. No subscription, nothing to cancel.
Open source models only. Offer ends June 30.
Unlimited open source LLM inference. Every open source model. Zero cost.
Open source models caught up where it matters for your app.
For chat, companions, tutoring, roleplay and summarisation, the output quality now matches the closed source models, but at a fraction of the price. Closed source still wins the hardest reasoning and agent orchestration. Your app doesn't need that.
What your app does need is a gross margin. When users chat all day against thin consumer pricing, inference is the line that decides whether the business works. The gap between closed list price and open source models on our hardware is 10 to 50x. This isn't a one-off discount, it's a structural pricing difference, and at consumer scale it's what separates an app that makes money from one that loses it on every conversation.
You already buy inference somewhere. We're asking you to point it at us for two weeks, at zero cost, and look at the numbers.
Lower cost than the frontier closed model you'd otherwise run.
The gap between closed list price and open source on our hardware.
Unlimited use of every open source model, at zero cost.
What 5,000 conversations cost.
A typical session: 30 turns, about 2K input tokens per turn with context, 200 output tokens per turn. Here's the bill on DeepSeek-V4-Flash next to a GPT-class closed model at list price.
Session model: 30 turns, 2K input plus 200 output tokens per turn. Closed comparison uses GPT-class list pricing. Same conversation quality. The difference is the bill.
Every open source model, unlimited for 14 days.
Not a starter subset. The whole catalogue, including the expensive ones. Qwen by Alibaba, GLM by Z.ai, Kimi by Moonshot AI, plus DeepSeek, MiniMax and Gemma by Google.
| Model | Model ID | In $/M | Out $/M | Good for | Open in playground |
|---|---|---|---|---|---|
| DeepSeek-V4-Flash | deepseek:v4@flash | $0.14 | $0.28 | Cheapest high-volume workhorse, 1M context, tool use | Playground |
| MiniMax M2.5 | minimax:m2.5@0 | $0.27 | $0.95 | Cheap, strong general workhorse | Playground |
| MiniMax M2.7 | minimax:m2.7@0 | $0.30 | $1.20 | Higher-capability tier | Playground |
| MiniMax M2.7 Highspeed | minimax:m2.7@highspeed | $0.60 | $2.40 | Latency-tuned | Playground |
| Qwen3.5-27B | alibaba:qwen3.5@27b | $0.24 | $2.00 | 262K context, multilingual workhorse | Playground |
| Qwen3.5-397B | alibaba:qwen3.5@397b | $0.50 | $3.20 | Long-context reasoning and memory, 262K to ~1M | Playground |
| Kimi K2.6 | moonshotai:kimi-k2.6 | $0.90 | $4.20 | Multimodal, agents | Playground |
| GLM-4.7 | zai:[email protected] | $0.40 | $1.75 | Strong generalist | Playground |
| GLM-5.1 | zai:[email protected] | $1.40 | $4.40 | 200K context | Playground |
| Gemma 4 31B | google:gemma@4-31b | $0.12 | $0.37 | Small, cheap, multilingual | Playground |
Prices shown are short-context rates. Long-context tiers vary by model.
Pick your app. Start with these models.
Each saving is the monthly difference versus the frontier closed model you'd otherwise use, on the same session model as the calculator above.
Companion & character chat
The heaviest LLM use there is: long, multi-turn sessions where cost scales straight with engagement. Your model price, plus backup capacity for when a provider goes down, decides whether the core loop holds.
Chat assistants & consumer apps
You monetise generic chat, homework and productivity at consumer scale: millions of light but frequent sessions. Margin lives in the blended cost per query, and you're probably already routing across models.
Language tutors
Conversational practice at scale, where cost per conversation, voice and chat latency, and steady multilingual quality move retention and margin directly.
Wellness & coaching
Long, recurring sessions with higher safety and compliance stakes. Cost matters, but trust comes first: no training on your data, zero retention, GDPR.
Your code already works here.
The endpoint is OpenAI-compatible. Your SDK, request shape and streaming format carry over unchanged. Reasoning models and SSE streaming work the way you expect. There's no rewrite hiding behind step 3.
from openai import OpenAI
client = OpenAI(
api_key="your_runware_api_key",
base_url="https://api.runware.ai/v1",
)
response = client.chat.completions.create(
model="deepseek:v4@flash",
messages=[{"role": "user", "content": "Hi"}],
stream=True,
)Lowest cost because we built it, not because we discounted it.
Open source models on Runware run on hardware we designed ourselves: the Sonic Inference Engine, tuned end to end for throughput. We're not reselling someone else's capacity or routing your requests through a third provider, which is why the prices hold and the latency doesn't wander. p50 and p95 stay flat when your traffic doubles.
It also makes us a real backup. When your current provider has an outage, rate-limits you, or reprices, your core loop keeps running.
More about the Runware inference stackBenchmark figures pending final validation.
Not just open source. Not just chat.
The same account also runs the latest frontier closed source LLMs at pass-through rates, plus image, video, audio and 3D through the native API. If your app generates avatars or scenes alongside chat, it's one integration and one invoice.
On day 15 your account rolls to standard pay-as-you-go open source rates. The prices in the table above, no markup, no subscription, no minimum. If the numbers didn't convince you, you walk away owing nothing.
FAQ
Your 14 days start when you do.
Activate before June 30. Two weeks of unlimited open source inference, then pay-as-you-go if you stay.