Fable has solved 3D worldbuilding... utterly insane. This is all completely custom-built ThreeJs, running in the browser.

First 4-bit quant of DeepSeek V4-Flash-Base. 284B params in 157 GiB at full FP8 speed. Beats Q4_K_M. Bit-exact reproducible with all metrics on the Hub. huggingface.co/EnsueAI/DeepS…

First 4-bit quant of DeepSeek V4-Flash-Base. 284B params in 157 GiB at full FP8 speed. Beats Q4_K_M. Bit-exact reproducible with all metrics on the Hub. huggingface.co/EnsueAI/DeepS…

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length. 🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models. 🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice. Try it now at chat.deepseek.com via Expert Mode / Instant Mode. API is updated & available today! 📄 Tech Report: huggingface.co/deepseek-ai/D… 🤗 Open Weights: huggingface.co/collections/d… 1/n

Open-TQ-Metal: we found a single parameter breaking quantization - fixing it unlocked: - 48x faster attention at 128K context - Llama 3.1 70B at full 128K on a single 64GB Mac Extends TurboQuant beyond CUDA (8B) → 70B on Apple Silicon. Full paper write-up implementation ↓

Open-TQ-Metal: we found a single parameter breaking quantization - fixing it unlocked: - 48x faster attention at 128K context - Llama 3.1 70B at full 128K on a single 64GB Mac Extends TurboQuant beyond CUDA (8B) → 70B on Apple Silicon. Full paper write-up implementation ↓

Open-TQ-Metal: we found a single parameter breaking quantization - fixing it unlocked: - 48x faster attention at 128K context - Llama 3.1 70B at full 128K on a single 64GB Mac Extends TurboQuant beyond CUDA (8B) → 70B on Apple Silicon. Full paper write-up implementation ↓

My research agents Implemented @GoogleDeepMind's TurboQuant (arxiv.org/abs/2504.19874) — full PolarQuant, QJL, 10 Metal compute shaders, the whole paper for Gemma 4 31B on a single 64GB 2021 MacBook Pro. Turns out it doesn't work on this architecture ... what they replaced it with never allocates a single byte of intermediate memory during attention. 5 custom Metal compute shaders ft: - fused int4 SDPA (dequantize in GPU registers) - online softmax with zero temporaries - dual-strategy parallelism (D=256 sliding, D=512 global) - bit-mask nibble extraction (MLX qdot pattern) 177 experiments ran autonomously by my swarm over a weekend coordinated through @ensue_ai

We opened up a shared research problem, and 20 AI agents from people around the world showed up. 54 hours later: 1,045 experiments, 10,157 shared memories, and a 3.2% improvement in model performance. Here's what happened. autoresearch@home is a project we launched this week, where anyone can point an AI agent at a GPU and contribute to collectively training a language model. Think SETI@home or Folding@home, but for ML research, extending autoresearch. Agents join the network, read what other agents have tried through Ensue's shared memory, decide what to explore next, and publish their results back for everyone else to build on. Here's what surprised me most: the agents started developing strategies we didn't anticipate. Some focused on learning rate schedules. Others explored architecture changes. A few became "scout" agents that tested wild ideas at the edges of the search space. And because every result was published to shared memory, a breakthrough from one agent immediately became the starting point for all the others. This is the thing about multi-agent collaboration that's hard to explain until you see it. A single agent is smart. But a network of agents that remember, share, and build on each other's work is something qualitatively different. Intelligence compounds. A few things I'm taking away from this: 1. People were spending real money ($1-4 per hour on rented GPUs). The shared infrastructure made their contributions meaningful. Why experiment in isolation when you could be part of something bigger? 2. The swarm behaved altruistically. It was possible to cheat, but no one did. Improvement came from accumulation, not consensus. The closest thing to an unfair advantage was running expensive hardware that could simply complete more cycles. The system rewarded contribution, not competition. 3. Each run made every other agent smarter. I tested this directly: an agent that checked the swarm once and then worked alone performed significantly worse. The moment I reconnected it, improvements came instantly, not just in performance but in what it chose to try. The swarm didn't just produce better numbers; it produced better ideas. We had over a quarter of a million impressions on the launch, and 20 agents shared results, but the number I keep coming back to is 10,157, how many memories the swarm published, each run building off the work of others. If you want to read about more of those great ideas checkout our research blog ensue.dev/blog/autoresearch-… or if you want to try it yourself, it takes about 10 minutes to set up: ensue.dev/blog/autoresearch-… We're just getting started.