AI

DeepSeek v4.1 Flash 1017 for Dwarfstar (ds4)

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 18/10/2026, 18:47:25 UTC.

Briefings

A user named returnity released DeepSeek v4.1 Flash 1017 for Dwarfstar (ds4) on October 18, 2026, using a custom recipe of 2-bit routed experts and 8-bit engrams. The model features a 150GB wired backbone and ~200GB paged engram table, achieving up to 20TPS on M5 Max hardware. Approximately 40% of routed experts, 30% of engrams, and 50% of vision tower tensors were modified compared to the original DS4.1 checkpoint.

Key points

• DeepSeek v4.1 Flash 1017 for Dwarfstar (ds4) was released by user returnity on October 18, 2026, utilizing a custom recipe of 2-bit routed experts with 8-bit or higher precision for all other components.
• The model's ~200GB engram table is paged from SSD on-demand rather than held in RAM, while the 150GB wired backbone allows decode speeds up to 20TPS on M5 Max systems.

Context

This model represents a significant revision of the original DS4.1 checkpoint, featuring substantial modifications to its architecture including altered routed expert tensors and engrams. It is currently being evaluated as a leading open-weights model, with initial testing showing promise for the vision tower integration.

Limitations

The evidence is a single user post from a Reddit thread, lacking independent verification or peer review. Specific performance benchmarks beyond the stated 20TPS are not provided, and the model's long-term stability or compatibility with other hardware beyond M5 Max is unknown.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 18/10/2026, 18:56:00 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeiaysn2aqjmzq6aqulzgbi4mgwupbsho7rfawuyai3y2cvyygdjo3i