Skip to main content
NeuronFeed
CATEGORY

Best Open Source AI AI Tools

35 tools compared · 2026

Llama 4, Qwen 3.5, DeepSeek V4, Mistral Large 3 — the open-weight stack that closed the gap in 2026

35 open source ai startups tracked, with the largest concentration in US. Total tracked funding: $12.8B.

Tracked
35
Total Raised
$12.8B
Countries
9
Active Deals
3

Editor's picks

6

Top by score

View all 35 →

Funding by year — Open Source AI

2022 → 2026
$2.3M
’22
$1.3B
’23
$2.6B
’24
$7.2B
’25
$1.1B
’26

Market overview

By April 29, 2026, five frontier-class open-weight LLMs had shipped within thirty days of each other: Meta's Llama 4 (Scout and Maverick), Alibaba's Qwen 3.5, DeepSeek V4 (Pro and Flash), Google's Gemma 4, and Mistral Medium 3.5. Almost every one of them is a sparse Mixture-of-Experts — DeepSeek V4-Pro at 1.6T total parameters with 49B active, Llama 4 Maverick at 400B/17B, Qwen 3.5 at 397B/17B, Mistral Large 3 at 675B/41B. The benchmark gap between open and closed is no longer the story; the licensing fragmentation is.

Apache 2.0 has effectively won the permissive license war among labs that want enterprise adoption. Gemma 4, Qwen 3.5, Mistral Large 3, and Yi all ship under it. DeepSeek V4 ships MIT, which is broader still. Llama 4 retains the Meta community license with use-based restrictions, which is why some commercial deployers prefer the Apache cohort even when Llama benchmarks higher on Llama 4 Maverick's 85.5% MMLU.

The 20 companies we track here account for roughly $8.3B in disclosed funding. Mistral AI ($6.34B cumulative including a recent debt facility, $13.8B post-money, Paris) is the headline European bet. Hugging Face ($470M Series D, $4.5B) is the registry every other lab depends on — Llama 4, Qwen 3.5, DeepSeek V4, Gemma 4, and Mistral Large 3 all distribute through the Hub. Liquid AI ($300M Series B) ships its LFM family as openly licensed checkpoints aimed at on-device deployment. Physical Intelligence ($735M Series B at $2.7B) extends the open ethos into robotics foundation models. Imbue ($232M Series B at $1B) builds open-research coding agents. Qdrant ($85M Series B, Berlin), LanceDB ($41.5M Series A, $155M), and Chroma ($20.3M seed at $75M) anchor the open-core retrieval layer that every RAG stack imports.

On the smaller end, Cline ($4M seed at $110M) is the open-source coding agent now scoring competitively against Cursor on Aider Polyglot when paired with GLM-5 or DeepSeek V3.2-Speciale. OpenCode runs on zero raised. Aider, Tabby, Refact.ai (Netherlands), and Kilo Code fill out the self-hosted developer-tooling tier. Falcon LLM out of TII Abu Dhabi and Kyutai in Paris are publicly funded open labs working outside the Mistral and Hugging Face axis.

Twelve of twenty companies are US-headquartered; the rest split across France (Mistral, Kyutai), Germany (Qdrant), Netherlands (Refact.ai), China (Baichuan), and the UAE (Falcon). EU and Gulf sovereigns are actively funding open-weight alternatives so critical workloads do not depend on US APIs, with US GPU export controls and the EU AI Act's general-purpose thresholds shaping the next wave of releases. Andreessen Horowitz, Lightspeed, General Catalyst, and Sequoia lead deal count, often alongside Nvidia and the major clouds.

Key trends 2026

  • Apache 2.0 won the license fight. Gemma 4, Qwen 3.5, Mistral Large 3, and Yi all ship Apache; DeepSeek V4 is MIT. Llama 4's community license with use restrictions is now the outlier, which shapes which checkpoints enterprises actually deploy.
  • Sparse MoE became the default architecture. DeepSeek V4-Pro (1.6T/49B), Llama 4 Maverick (400B/17B), Qwen 3.5 (397B/17B), and Mistral Large 3 (675B/41B) all ship as MoE — a structural choice that bends the cost curve favorably for inference providers like Together AI and Fireworks.
  • China's open-weight surge is the pricing pressure. Qwen 3.5, DeepSeek V4, and GLM-5 hit benchmark parity with closed US models at a fraction of the per-token cost, which is part of why Anthropic and OpenAI margins keep getting questioned.
  • EU and Gulf sovereignty money is building a parallel stack. Mistral's $13.8B mark and Falcon LLM's TII funding are explicit bets that critical workloads should not depend on US closed APIs — a thesis the EU AI Act actively reinforces.

Benchmarks vs global

Tracked open-source AI companies
20
foundation labs + OSS infra
Combined disclosed funding
$8.3B
Mistral = ~76% of total
Largest single raise
$6.34B
Mistral cumulative incl. debt
Best open-weight SWE-bench
77.8%
GLM-5 vs Claude Opus 4.7 at 87.6%

Top countries

By startup count

Stage breakdown

Latest round type
  • Seed 10
  • Series A 9
  • Series B 4
  • Series C 2
  • Series E 1
  • Grant 1
  • Debt 1

Top investors backing Open Source AI

See all →

FAQ

Frequently asked

Which open-weight model leads on coding benchmarks in 2026?
GLM-5 hit 77.8% on SWE-bench Verified — the strongest open-weight result publicly reported. DeepSeek V3.2-Speciale and Qwen3-Coder are close behind. Closed leaders Claude Opus 4.7 (87.6% SWE-bench Verified) and Claude Opus 4.5 (89.4% Aider Polyglot) still lead, but the gap is roughly 10 points, not a generation.
Are Hugging Face and Mistral really comparable companies?
They are very different bets. Hugging Face ($470M Series D, $4.5B) monetizes the registry — Llama 4, Qwen 3.5, DeepSeek V4, and Gemma 4 all flow through the Hub. Mistral AI ($6.34B cumulative, $13.8B) is a foundation lab competing directly with closed labs while shipping open weights as part of its commercial strategy.
How does US export control policy affect this category?
Recent rules limit advanced GPU shipments to certain jurisdictions and create reporting thresholds for very large training runs. US-based labs that want to release weights freely now navigate those thresholds carefully. The pressure has accelerated EU sovereignty deals, with Mistral as the headline beneficiary and Falcon LLM at TII in Abu Dhabi pursuing a similar mandate.
Why are Chinese open-weight models suddenly the pricing benchmark?
DeepSeek V4, Qwen 3.5, and GLM-5 closed most of the benchmark gap to closed US models while remaining permissively licensed. Inference providers like Together AI and Fireworks AI host them at a fraction of OpenAI or Anthropic per-token rates — Fireworks' blended rate sits near $0.84 per 1M tokens. That changes the build-versus-fork calculus for every downstream developer.
Which open-source coding agents matter in 2026?
Cline ($4M seed at $110M) runs as a VS Code extension with bring-your-own-model support, often pointed at GLM-5 or DeepSeek V3.2-Speciale. OpenCode is the terminal-side Aider alternative. Aider itself remains a reference implementation. Imbue's $232M Series B funds research-heavy collaborative agents. Refact.ai (Netherlands) and Tabby anchor the self-hosted tier.

Recent rounds in Open Source AI

All rounds →
Date Startup Round Amount
May 2026 RadixArk Seed $100M
Mar 2026 Mistral AI Debt $830M
Mar 2026 Qdrant Series B $50M
Jan 2026 Inferact Seed $150M
Dec 2025 Kilo Code Seed $8M
Nov 2025 Physical Intelligence Series B $600M
Nov 2025 Numerai Series C $30M
Oct 2025 Reflection AI Series B $2B

All Open Source AI startups

Page 1

Mistral AI

Verified
FR est. 2023

Open and efficient foundation models

Raised
$6.3B
Stage
Debt
90

Reflection AI

US est. 2024

America's open frontier AI lab building autonomous coding agents

Raised
$2.1B
Stage
S-B
82

Baichuan AI

CN est. 2023

Empowering the world with knowledge and professional services through breakthrough language AI.

Raised
$1.0B
Stage
S-A
80

Zyphra

United States est. 2020

Open-science AI lab building efficient multimodal models and the Maia superagent

Raised
$110M
Stage
S-A
75

Inferact

US est. 2025

Commercial inference engine from the creators of vLLM

Raised
$150M
Stage
Seed
73

Kyutai

FR

An open-science AI lab dedicated to building and democratizing Artificial General Intelligence through open research.

Raised
$330M
Stage
Grant
72

RadixArk

US est. 2025

Open infrastructure for frontier AI, built on SGLang

Raised
$100M
Stage
Seed
72

Prime Intellect

US est. 2023

The open superintelligence stack for distributed, decentralized AI training

Raised
$70M
Stage
S-A
71

DeepInfra

US est. 2022

High-throughput, low-cost AI inference cloud

Raised
$133M
Stage
S-A
71

Hugging Face

US est. 2016

The AI community building the future of machine learning.

Raised
$970M
Stage
S-E
70

BentoML

US est. 2019

Unified inference platform to run AI at scale

Raised
$20M
Stage
Seed
70

Physical Intelligence

Verified
US est. 2023

Foundation models for physical world AI

Raised
$600M
Stage
S-B
68

Numerai

US est. 2015

The hardest data science tournament on the planet

Raised
$30M
Stage
S-C
66

TensorZero

US est. 2024

Open-source stack for building industrial-grade LLM applications

Raised
$7.3M
Stage
Seed
66

Hyperbolic

US est. 2022

Open-access AI cloud for compute and inference

Raised
$20M
Stage
Seed
66

Liquid AI

US est. 2023

Efficient general-purpose foundation models built for edge, on-device, and cloud deployment at every scale.

Raised
$250M
Stage
S-A
64

Sea AI Lab - SAIL

SG

Incubating the future by understanding and creating intelligence from first principles to solve challenging problems.

64

Unitree Robotics

PRIVATE
China est. 2016

Affordable quadruped and humanoid robots powered by embodied AI

Stage
S-C
64

Langfuse

DE est. 2023

Open source LLM engineering platform for debugging, analyzing, and iterating on LLM applications

Raised
$8M
Stage
Seed
63

Refact.ai

NL est. 2022

Self-hosted AI coding assistant — run on your own infra, own your data

Stage
Seed
60

Tabby

An open-source, self-hosted AI coding assistant for secure and efficient development.

60

OpenCode

US est. 2024

Open-source terminal AI coding agent — Aider alternative with multi-model support

59

Aider

US est. 2023

AI pair programming in your terminal that edits your codebase and commits changes via Git.

59

Nous Research

US est. 2023

Open-source AI lab building decentralized models, agents, and training

Raised
$65M
Stage
S-A
59