The AI map · Models and assistants

Open-weight models

Models whose weights you can download and run on your own servers, or call cheaply through a host, so your data and costs stay in your hands. The ones that lead it as of 29 October 2026.

  1. DeepSeek V4 (DeepSeek-V4-Pro and DeepSeek-V4.1-Flash) New since July

    DeepSeek

    Strong open models for coding and agents under the MIT licence, with a very cheap hosted API and a 1M-token context.

    Pick it You want near-frontier coding and agent work at the lowest API price, or MIT-licensed weights.

    Price Weights free (MIT). API per million tokens at peak: Flash $0.30 input / $1.20 output, V4-Pro $1.32 / $3.96; half price off-peak

    Source: api-docs.deepseek.com

  2. Qwen3.8 New since July

    Alibaba (Qwen)

    Alibaba's open family, from a 27B model that sees images and video to a very large flagship, for coding, research and agents.

    Pick it You need a capable model small enough to run yourself; Qwen3.8-27B is Apache 2.0.

    Price Weights free to download; Qwen3.8-27B is Apache 2.0, larger models carry Qwen's own licences

    Source: huggingface.co

  3. Kimi K3 New since July

    Moonshot AI

    A 2.8-trillion-parameter open model for long coding sessions and knowledge work, with vision and a 1M-token context.

    Pick it You want the largest open model for long agentic work and can host it or use the Kimi API.

    Price Weights free under the Kimi K3 License

    Source: kimi.com

  4. GLM-5.3 New since July

    Z.ai

    An open model aimed at complex coding and long-horizon tasks; Z.ai calls it the most capable open-weights model for coding.

    Pick it Coding is the main job; GLM-5.3-Flash is a smaller MIT-licensed option.

    Price Weights free; GLM-5.3 under its own licence, GLM-5.3-Flash under MIT

    Source: huggingface.co

  5. Gemma 4

    Google

    Small open models (2B to 31B) that run on a phone, a laptop or one GPU, and handle images, video and audio.

    Pick it You need a model on a device or a single machine rather than a data centre.

    Price Free to download

    Source: ai.google.dev

  6. gpt-oss-120b and gpt-oss-20b

    OpenAI

    OpenAI's open-weight reasoning models; the 120b fits on one 80GB GPU and the 20b suits local use.

    Pick it You want an Apache 2.0 model from OpenAI that runs on one GPU; it dates from August 2025, so newer open models score higher.

    Price Free to download (Apache 2.0)

    Source: huggingface.co

  7. Sarvam-105B and Sarvam-30B

    Sarvam AI

    Indian-built open models trained with a focus on the Indian context, with state-of-the-art results for their size across 22 Indian languages; they power Sarvam's Indus chat app.

    Pick it Your users write or speak in Indian languages and you want an Indian, Apache 2.0 model you can host yourself.

    Price Free to download (Apache 2.0)

    In India Built in India for 22 Indian languages; Indus, the chat app on Sarvam-105B, opened in limited beta in February 2026.

    Source: huggingface.co

What changed

  1. 17 Oct 2026 DeepSeek released DeepSeek-V4.1-Flash-1017, which now serves the deepseek-flash API name; the older V4-Flash models were retired from the API. api-docs.deepseek.com
  2. 6 Oct 2026 Mistral opened a preview of Mistral Large 4, a 1-trillion-parameter model, and said the weights would follow by the end of October. mistral.ai
  3. Sep 2026 DeepSeek released DeepSeek-V4.1-Flash. huggingface.co
  4. 13 Aug 2026 DeepSeek released DeepSeek-V4-Pro-0813, the official V4-Pro, under the MIT licence. huggingface.co
  5. Aug 2026 Alibaba released the Qwen3.8 open family, including Qwen3.8-27B under Apache 2.0. huggingface.co
  6. Aug 2026 Z.ai released GLM-5.3 and GLM-5.3-Flash. huggingface.co
  7. 31 Jul 2026 DeepSeek published DeepSeek-V4-Flash-0731. huggingface.co
  8. Jul 2026 Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model it calls the first open 3T-class model. kimi.com

Questions people ask about it

AI in Your Product

  1. Should my startup add AI features, and which ones actually matter to users?
  2. How do founders ship their first AI feature (APIs, wrappers, RAG in plain terms)?
  3. How much does it cost to run AI features, and how do I keep costs sane?
  4. Which AI model should I build on (GPT, Claude, Gemini, open-source), and how do I avoid locking myself in?
  5. How do I test whether my AI feature actually works before launch (evals, in plain terms)?
  6. How do I stop my AI feature from making things up in front of customers?

Building Your AI Stack

  1. ChatGPT vs Claude vs Gemini vs Grok, how do founders actually choose?
  2. Which paid AI subscriptions are worth it for an early-stage founder?
  3. How do I evaluate the endless stream of new AI tools without drowning?

The rest of the map

Checked 29 October 2026. Prices and rankings move every few months; each tool links the page its claims were read from.

eChai Partner Brands The partnership →