One pipeline, left to right: ingestion, chunking, embeddings, vector databases, routing, models, inference, deployment, guardrails and security, with the tools for every step.

Explore the map →

Latest analysis

All posts

Open Source Gems · Nov 7, 2026

Cerebras vs Groq vs SambaNova: where to run open models fast, and what it costs

Three companies built their own chips to sell open-model tokens faster than GPU clouds. Measured speed and price on the same models, the rate limits and data terms that matter in production, where each one wins, when Together AI or Fireworks AI is the better buy, and how much company risk you take on.

AI for Security & Compliance

Guardrails for AI apps and agents: what they catch, what they cost, and what matters more

Coding

Testing AI apps, and letting AI test yours: the layers, the tools and what they cost

Agent Frameworks & Automation

10 lesser-known MCP projects worth a look

Open source

Running giant AI models on an ordinary PC: Colibri, Strata, AirLLM and FreeToken compared

Just reviewed

More recent reviews

Harrier-OSS-v1 (270M, 0.6B, 27B)

MicrosoftEmbeddings & rerankingOpen weights

4.5/5

Microsoft's MIT-licensed text embedders in three sizes; the 27B led multilingual MTEB v2 at 74.3 and the small ones lead their classes.

Lean LSP MCP

Oliver DresslerCode & dev tools

4.5/5

Gives an agent Lean 4's proof state, errors and theorem search through the language server, so it can write proofs the checker accepts.

Nemotron 3.5 Embed (1B and 8B)

NVIDIAEmbeddings & rerankingOpen weights

4.4/5

NVIDIA's open embedders for text and page images in one vector space, first on ViDoRe v3.1 and free to call on OpenRouter.

ACE-Step 1.5

ACE Studio and StepFunMusic & audioOpen weights

4.3/5

MIT-licensed song model that runs in under 4 GB of VRAM, makes up to 10 minutes with lyrics in 50+ languages, and edits sections.

Eleven Music v2.5

ElevenLabsMusic & audio

4.3/5

ElevenLabs' music model: songs with vocals up to five minutes, section editing and wide commercial rights, at $0.15 a minute.

Qwen3-ASR (1.7B and 0.6B)

Alibaba (Qwen)Speech to textOpen weights

4.3/5

Alibaba's Apache 2.0 transcribers in 1.7B and 0.6B sizes: 30 languages, 22 Chinese dialects, language ID and streaming in one model.

The best-reviewed tools

All tools

ElevenLabs

Audio & voiceFree plan

4.8/5

AI voice platform for text to speech, voice cloning, dubbing, music and voice agents, now on its Eleven v4 models.

ChatGPT

Chat assistantsFree plan

4.7/5

OpenAI's general-purpose assistant, now on GPT-6, with chat, deep research, ChatGPT Work and Codex in one app.

Claude Code

CodingPaid

4.7/5

Anthropic's coding agent that reads your repo, edits files and runs commands, in the terminal, IDE, desktop app or browser.

Hugging Face

Model APIs & hostingFree plan

4.7/5

The main hub for open AI models and datasets, with app hosting, a multi-provider inference API and dedicated endpoints.

Intercom Fin

Customer supportPay per use

4.7/5

AI customer agent that answers and resolves support, sales and ecommerce conversations, billed $0.99 per outcome.

Canva

Design & slidesFree plan

4.6/5

Drag-and-drop design platform for social posts, presentations, video and docs, with Canva AI 2.0, Magic Layers and free Affinity.

Models, ranked

All models

Top-rated AI models
Model	Made by	Type	Open weights	Rating
Claude Opus 5.5	Anthropic	General purpose	No	4.7/5
Eleven v4 and v4 Turbo	ElevenLabs	Text to speech	No	4.7/5
GPT Image 2.5 (Flare and Sunburst)	OpenAI	Image generation	No	4.7/5
Claude Sonnet 5.5	Anthropic	General purpose	No	4.6/5
GPT-6.1 Sol	OpenAI	General purpose	No	4.6/5
Nano Banana 2.1	Google	Image generation	No	4.6/5
Voyage rerank-3	Voyage AI (MongoDB)	Embeddings & reranking	No	4.6/5
Claude Haiku 5.5	Anthropic	General purpose	No	4.5/5

MCP servers worth installing

All MCP servers

MCP servers let assistants such as Claude, ChatGPT and Cursor reach into other apps and data. These are the ones that earn a place in your setup, and the permissions each one asks for.