The system relies on a compress-then-correct mechanism created by Ryan Flavell that uses a 2-bit copy of model experts alongside stored correction planes. This technical approach achieved 98.3 percent output parity with the full model during a 960-word audit. It delivers speeds reaching 7.4 tokens per second on warm-cache code prompts while operating under an Apache-2.0 license. External tools can integrate directly with the engine via an OpenAI-compatible local endpoint that accommodates streaming.
Key points
Users can run massive AI models privately on consumer workstations without transmitting s…: Users can run massive AI models privately on consumer workstations without transmitting sensitive conversational data to remote servers.
The engine matches 98.3 percent of the original architecture's output fidelity while mini…: The engine matches 98.3 percent of the original architecture's output fidelity while minimizing memory pressure through selective weight reconstruction.
Developers can immediately integrate the local model into existing workflows through an O…: Developers can immediately integrate the local model into existing workflows through an OpenAI-compatible endpoint supporting streaming and tool calls.
Sources checked by our analysis pipeline: 1
The original model, not a smaller copy · 64 GB of RAM and one 12 GB NVIDIA card · free and open source
PROMETHEVS runs DeepSeek-V4.1-Flash-1017, one of the strongest open models, from all 510 GB of its original weights, on an ordinary PC that cannot even load it: a Ryzen 9 3950X with 64 GB of DDR4, an RTX 3060 with 12 GB and two NVMe drives.
The PROMETHEVS engine gets there with compress-then-correct, a method devised by Ryan Flavell. Every word is first computed on a 2-bit copy of the model's experts. When the model itself is unsure (its top two choices are close), that word is recomputed with the original weights, rebuilt exactly from a stored correction plane. You get close to the speed of the compressed copy and close to the answers of the full model.
- Full-model quality, counted. On 960 audited words the output picks the full model's word 98.3% of the time (the 2-bit copy alone: 97.4%).
- Up to 7.4 tokens/s on a code prompt (warm cache, best case), against 2.14 for…
Microsoft is winding down an agreement with Oracle to rent artificial intelligence cloud capacity, according to a report from Bloomberg News. Details regarding the timeline or reasons for concluding the arrangement between the two technology companies were not disclosed in the report.
Federal prosecutors are pursuing charges against three individuals from abroad who allegedly leveraged artificial intelligence to mimic high-profile corporate leaders and legal counsel. The sophisticated impersonation operation targeted victims to execute illicit transfers exceeding $129 million. This federal action highlights the growing legal and financial perils posed by synthetic media deployed against corporate executives.
Recent midterm election results across the United States have ushered in Democratic governors and federal lawmakers poised to heighten oversight on the technology sector. Local resistance to critical computing infrastructure also surged, with voters overwhelmingly opting to curb development projects nationwide. This shifting political landscape poses mounting regulatory challenges for companies expanding operations to support advanced systems.