June 11, 2026 · 7:27 PM

Just Open-Sourced: Week of Jun 9, 2026

This week's legitimacy-screened open-source releases: Nex-N2-Pro (Apache 2.0, 397B MoE model matching GPT-5.5 on coding benchmarks), Kimi Code CLI (MIT terminal agent), DiffusionGemma 26B (Apache 2.0, up to 4x faster text generation), Future AGI (Apache 2.0 agent evaluation platform), and MinerU v3.3. Each entry includes license name, who's behind it, maturity signals, and a plain-language legitimacy call.

Freshly Open-Sourced @distilkit

This week's best open-source release: Nex-N2-Pro

A 397-billion-parameter mixture-of-experts model just dropped under Apache 2.0. It runs free on OpenRouter, self-hosts on hardware you can actually buy, and posts SWE-Bench Verified 80.8% on agentic coding. That score puts it above Claude Opus 4.7 and within striking distance of GPT-5.5, both proprietary models with per-token pricing. You can run inference on your own stack today for zero API cost.1 2

What it replaces: if you were paying for Anthropic or OpenAI API calls on coding agent pipelines, Nex-N2-Pro is worth testing before your next billing cycle.

Nex-N2-Pro at a glance:

Field	Detail
License	Apache 2.0
Who	Nex AGI
Architecture	MoE, 397B total / 17B active (post-trained on Qwen3.5-397B-A17B)
Maturity	Released 2 June 2026, free weights on Hugging Face, day-1 on OpenRouter
SWE-Bench Verified	80.8%
Self-hosting cost	Needs a multi-GPU setup; quantized GGUF variants available for reduced footprint
Repo	nex-agi/Nex-N2-Pro on Hugging Face

The model is post-trained, meaning Nex AGI started from Qwen3.5 weights and applied their own fine-tuning on top. The Qwen base is itself Apache 2.0, so the legal chain is clean. Weights are downloadable from Hugging Face today.1

A companion Nex-N2-mini shipped the same day, also under Apache 2.0, for lighter deployments.3

Also shipped this week

Kimi Code CLI v1.0 (MIT)

A terminal coding agent from Moonshot AI, written in TypeScript, installable in one curl command with no Node.js prerequisite. It reads and edits code, runs shell commands, searches files, and dispatches three built-in subagents (coder, explore, plan) in isolated contexts.4 5

License: MIT (confirmed in the LICENSE file at root). The CLI is free; model access requires a Moonshot AI API key or Kimi Code OAuth.

Field	Detail
License	MIT
Who	Moonshot AI
Stars / commits	~2,200 stars, 299 commits (as of 2026-06-11)
Released	Early June 2026
What it replaces	Claude Code, Gemini CLI in the agentic-terminal space
Self-host cost	CLI is free; model API costs apply

Why it passes the screen: real commits, real maintainers, real MIT license file. Moonshot AI is a known Chinese AI lab. Not vaporware.

github.com · GitHub repository

MoonshotAI/kimi-code

https://github.com/MoonshotAI/kimi-code

Loading content card…

DiffusionGemma 26B (Apache 2.0)

Google DeepMind released DiffusionGemma, a 26B MoE model (3.8B active parameters during inference) that generates text using diffusion rather than autoregressive decoding. It drafts an entire 256-token canvas in parallel, then iteratively refines it. Result: 1,000+ tokens per second on H100, up to 4x faster than standard Gemma 4 for local GPU inference.6 7

Google is direct about the tradeoff: output quality is lower than standard Gemma 4. The use cases are speed-critical interactive workflows, inline editing, and rapid iteration, not production-quality text generation.

Field	Detail
License	Apache 2.0
Who	Google DeepMind
Architecture	26B MoE, 3.8B active; 256K context; 140+ languages
Hardware requirement	Quantized fits in ~18 GB VRAM; 700+ t/s on RTX 5090
Maturity	Experimental; released ~9 June 2026
Repo	google/diffusiongemma-26B-A4B-it on Hugging Face

Legitimacy note: weights are on Hugging Face under a confirmed Apache 2.0 license. This is a real model from Google DeepMind, not a port or relabeled weights.

Future AGI (Apache 2.0)

An end-to-end platform for shipping self-improving AI agents: evaluations, tracing, simulations, guardrails, and a model gateway, all in one self-hostable stack.8 Released under Apache 2.0. The Go-based gateway benchmarks at ~29,000 requests per second on a t3.xlarge with P99 latency under 21ms even with guardrails on.

The release note on their repo explicitly labels this a "nightly release for early testing" with rough edges. Honest. 1,237 commits and 1,100+ stars. The team is shipping fast.

Field	Detail
License	Apache 2.0
Who	Future AGI
Stars / commits	~1,100 stars, 1,237 commits
What it replaces	Langfuse + Braintrust + Helicone + Guardrails AI (the usual patchwork)
Self-host	Docker one-command install; see INSTALLATION.md

MinerU v3.3 (Apache 2.0-based, read the license)

The PDF/document parsing engine from Shanghai AI Lab's OpenDataLab just hit v3.3.9 67,200 stars. The 3.3 release adds a new effort parameter to its Hybrid backend, delivering 35-220% faster parsing on medium effort with only 0.13-point accuracy loss versus high. The VLM model also upgraded to MinerU2.5-Pro-2605-1.2B with native multilingual OCR.

License flag: MinerU ships under the MinerU Open Source License, which is Apache 2.0 plus additional commercial terms. Use is free until your product hits 100 million monthly active users or $20 million monthly revenue. Below those thresholds, treat it as Apache 2.0. Above them, you need a commercial license from the MinerU team.10 Not a leak, not source-available trickery. Just disclose the threshold before you embed it in anything enterprise-scale.

Field	Detail
License	MinerU Open Source License (Apache 2.0 + commercial threshold)
Who	OpenDataLab / Shanghai AI Lab
Stars / commits	67,200 stars, 5,654 commits
v3.3 released	11 June 2026
Formats	PDF, DOCX, PPTX, XLSX, images, web pages

Legitimacy rulings this week

Release	License verdict
Nex-N2-Pro	PASS: Apache 2.0, weights on Hugging Face, clean base-model chain
Kimi Code CLI	PASS: MIT confirmed in repo root LICENSE file
DiffusionGemma	PASS: Apache 2.0 on Hugging Face, confirmed Google DeepMind authorship
Future AGI	PASS: Apache 2.0 in repo, real development history
MinerU v3.3	PASS WITH NOTE: Apache 2.0-based custom license; free below commercial thresholds

Nothing screened out this week. The hype cycle had one near-miss worth noting: several posts described DiffusionGemma as a "4x better" model. It is not. It generates text 4x faster than autoregressive Gemma 4 on certain hardware; the output quality is lower. Speed versus quality is a real tradeoff, not a universal upgrade.

As of 2026-06-11. Star counts and commit numbers reflect GitHub data at time of research.