A few days ago Anthropic shipped Claude Fable 5 - the first public cut of its Mythos class, the line it had been telling everyone was "too powerful to release." Three days later the US Commerce Department hit it with an export-control order, and Anthropic had to "abruptly disable" Fable 5 and Mythos 5 for every customer, everywhere. Not just foreign nationals - everyone, because that was the only way to comply.
Most of us were just poking at it. Running benchmarks, seeing whether it lived up to the Mythos framing, maybe wiring it into a side project to find out what it could actually do. Not a lot of people had a real product riding on Fable 5, so I'm not going to pretend the sky fell.
But that's the thing about an endpoint you don't control. It was there on Monday and gone by Thursday, and there was nothing any customer could do about it except wait. Rent your model from someone else and they can pull it whenever a regulator, a lawyer, or a launch plan says so. Weights you download and run yourself can't be revoked.
The gap basically closed
While Fable 5 was getting pulled, three open-weight models landed that you don't have to make excuses for. They run on hardware you already own, and they're close enough to the frontier that self-hosting isn't a quality sacrifice anymore.
Gemma 4 with QAT
Google shipped Gemma 4 QAT checkpoints on June 5. QAT is quantisation-aware training - you simulate the quantisation during training so the model keeps its quality when you compress it, rather than shrinking it afterwards and hoping. The numbers are real: roughly a 72% VRAM cut versus the BF16 baseline, the little E2B is down to about 1GB, and the 26B-A4B that wanted ~17GB now runs in ~15GB - which means it fits on a 16GB card.
Sizes run from E2B up to 31B, with a 26B-A4B mixture-of-experts in the middle. The little E2B runs anywhere there's a gigabyte going spare - a phone, a laptop, an edge box. The 26B-A4B is the sweet spot: about 15GB means it fits on a single 16GB card (an RTX 4080 or 4070 Ti Super) or any 16GB+ Mac. The full 31B wants more - a 24GB card or a 32GB+ Mac. That's QAT doing its job; a 26B model that fits in 16GB at all is the whole point.
Liquid LFM2.5
Liquid AI's LFM2.5-8B-A1B, from May 28. A sparse mixture-of-experts: 8.3B parameters total, but only 1.5B active per token, with a 128K context window and chain-of-thought reasoning. Only 1.5B firing per token means the compute is tiny, so this is the one that runs on almost anything - laptops, phones, a Mac Mini, even a Raspberry Pi if you're feeling brave. The pitch is fast, reliable tool calling on whatever's already on your desk, and for an on-device agent it's a fantastic fit.
Qwen 3.6 and the Qwopus merges
Qwen 3.6 arrived in April - hybrid Gated DeltaNet attention plus a sparse mixture-of-experts and a native vision encoder, with the whole line reoriented around agentic coding. Two open variants: a 27B dense and a 35B-A3B MoE (35B total, 3B active). The dense 27B wants a 24GB card (a 3090 or 4090) or a 32GB+ Mac. The 35B-A3B MoE still needs that same 24GB card to hold the weights, but only fires 3B per token - cheaper per token than the dense model, which matters for coding where you want throughput.
Then there's Qwopus. The community has been busy - Jackrong's Qwopus3.6-27B-v2 takes the 27B, rebuilds the reasoning traces, adds Multi-Token Prediction for speed, and ships as GGUF. It runs in llama.cpp and Ollama. When people say the local coding experience is finally good enough to actually use day to day, this is the model they mean.
A model on disk isn't an API
But here's the catch. A model on disk is not a service. To run these in production you need the boring layer - routing across machines, auth and keys, billing, failover when a box disappears, the OpenAI-shaped endpoint your tooling already speaks. That's the bit that's been missing.
That's what Infersec is for. You run our conduit agent on your hardware - Macs with Apple Silicon, NVIDIA rigs, AMD accelerators - and it dials home, pulls the model config, and serves. Point an endpoint at your sources and out comes an OpenAI- and Anthropic-compatible API. HuggingFace integration fetches the weights, MCP handles tool calls server-side, and the whole thing runs in the EU with no prompt logging.
Point your agents at it
Because the endpoint speaks OpenAI, everything you'd plug into a frontier model just works. opencode, openclaw, nanoclaw, hermes agent, mastra.ai, LangGraph - swap the base URL for your Infersec endpoint, drop in an API key, and you're running on weights nobody can take back.
Open weights got good enough to bet on. Fable 5 is a reminder of why you'd want to.
Sign up for early access, or see what we recommend running on the Recommended Models page.