Free models and new providers of the week (Nov 13)

!Cover art for the bridgecode weekly free models recap

Free and open source models

Runway Praxis-1 (11/11)

Robotics action model (world action model) with 27B parameters, MoE with diffusion transformer, predicting video and actions together. Weights, inference code, and fine-tuning code (full and LoRA) released under Apache 2.0, with commercial use permitted. Runway cites 46.2% success on 1,200 RoboLab-120 episodes, versus 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano, and 28.0% for π0.5. Runs on data center GPUs (H100, B200), workstation GPUs (RTX 5090, RTX PRO 6000), and Jetson Thor. License price: US$ 0 (R$ 0.00); hosted inference on Reactor had no price disclosed. Official source: https://runway.com/news/company-news/praxis-1-is-now-available

MiniMax H3-Image (11/10)

Open-weight image model with 33.1B parameters (DiT) for text-to-image, generation with up to 16 references, and editing, with native output up to 4096×4096. Weights are on Hugging Face under MiniMax's community license, and the model is also available via the company's API. MiniMax says it costs 50% to 70% of comparable offerings, with a 50% discount on 4K during launch week, but did not publish a per-image price in the consulted text. Source: https://cryptobriefing.com/minimax-h3-image-open-weight-4k-model/. Coverage reports that the license excludes the US, EU, UK, and South Korea; read the terms before use.

Thinking Machines Inkling 2 Small (11/12)

Launched on 11/05 with closed weights, gained public weights on Hugging Face under Apache 2.0 on 11/12, per AI Tracker, fulfilling the promise made at launch. Weights price: US$ 0 (R$ 0.00). Source: https://ai-tracker.rootstudio.io/news/2026-11-12-inkling-2-small-weights

Zyphra ZONOS3-Mini (11/12)

TTS with zero-shot voice cloning in 43 languages, MoE with 100M active parameters (800M total), weights Apache 2.0 and code MIT. Zyphra reports under 570 ms to first audio in a Pixel 9 browser and under 150 ms on a MacBook Air M3. Price: US$ 0 (R$ 0.00). Official source: https://huggingface.co/blog/Zyphra/zonos3-mini

Tether AI Research BotPsy (11/11)

0.8B model for function calling and on-device agents, based on Qwen3.5-0.8B. The announcement says Apache 2.0 weights "will be released," so download availability still needs confirmation. Official source: https://huggingface.co/blog/qvac/botpsy

Quick comparison

• Praxis-1: robotics, 27B, Apache 2.0, commercial use permitted.
• H3-Image: image, 33.1B, community license with reported geographic restrictions.
• Inkling 2 Small: text, Apache 2.0 since 11/12.
• ZONOS3-Mini: voice, 100M active, Apache 2.0, runs in browser.
• BotPsy: 0.8B agent, weights promised, still to confirm.

For developers, the license weighs as much as the benchmark. Apache 2.0 allows commercial use without asking permission, while community licenses often bring limits by region, revenue, or use type. Before putting any of these models into a product, read the license file in the repository and keep a copy of the version current on the day of download.

New providers and startups

• Qvyon Technologies (12/11): debuted with US$ 8 million (R$ 40.800.000,00) in funding led by MaC Venture Capital. Launched Mimic Core, a validation platform for autonomous vehicles and robots that transforms real sensor recordings into simulated scenarios. Source: https://siliconangle.com/2026/11/12/qvyon-launches-with-8m-to-improve-safety-for-physical-ai-deployments
• Clipto Studio (12/11): meeting and call memory layer for ChatGPT, Claude and other assistants via MCP, with local models on Mac and Windows. Source: https://siliconangle.com/2026/11/12/clipto-enables-chatgpt-and-claude-to-search-and-analyze-years-of-peoples-meetings-and-video-calls
• Concentrate AI (12/11): AI gateway for healthcare, with an API and a BAA for frontier and open-weight models. No public pricing in the announcement. Source: https://fortune.com/press-releases/concentrate-ai-hipaa-compliant-gateway-2026-11-12
• Phota Labs (12/11): iOS AI photo editing app, launched free (US$ 0 / R$ 0,00), with a US$ 5.6 million (R$ 28.560.000,00) seed led by a16z the previous year. Source: https://techcrunch.com/2026/11/12/phota-labs-uses-ai-to-turn-everyday-photos-into-professional-looking-shots
• Nectar (12/11): AI agent that finds creators among more than 300 million profiles and handles outreach to them, aimed at brands. Source: https://techcrunch.com/2026/11/12/nectar-launches-an-ai-agent-to-automate-creator-discovery-and-outreach-for-brands

How to test today

1. Download the weights for Praxis-1 or ZONOS3-Mini from the official repositories and run the README example. ZONOS3-Mini has a Space on Hugging Face and a WebGPU playground to test without installing anything.
2. For Inkling 2 Small, check on Hugging Face if the official repository already shows the Apache 2.0 license before downloading.
3. For H3-Image, use the API during launch week (50% discount on 4K, according to MiniMax) if you want to compare quality before hosting the weights.
4. Record the license, checkpoint version and download date for each test.

What to watch

• Open weights are migrating from text to robotics, voice and image, with different licenses between them.
• Launching closed and opening the weights a week later, as Thinking Machines did, has become a pattern to follow.
• The startup market remains driven by vertical agents and by validation infrastructure for physical AI.