AI

[Release] ISTA DeepSeek-V4.1-Flash-1017 GSQ-RCO Coder: a 763B model at 1 bit per parameter

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/11/2026, 16:47:26 UTC.

Briefings

The ISTA DeepSeek-V4.1-Flash-1017 model was released as a sparse Coder build with 1 bit per parameter, reducing transformer weights from 138.2 GB to 72.1 GB while maintaining 763.2B total parameters. This 763B model uses GSQ and RCO quantization to achieve near-optimal compression, with the Coder variant retaining 98.7% SWE-bench Verified performance compared to the base model.

Key points

• The model achieves 1 bit per parameter by removing 6 active experts per token from a 384-expert layer, reducing transformer weights from 138.2 GB to 72.1 GB.
• The Coder variant retains 98.7% SWE-bench Verified performance against the base model while maintaining 763.2B total parameters.

Context

This release represents a significant compression breakthrough for LLMs, enabling efficient deployment of a 763B model at 1 bit per parameter. The method combines GSQ and RCO quantization techniques to minimize memory usage while preserving task performance.

Limitations

Performance metrics are limited to benchmark scores on SWE-bench and LiveCodeBench. The Coder variant shows reduced performance on long agentic trajectories compared to single-problem code generation, and individual weights average 2.06 bits rather than 1 bit.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/11/2026, 07:25:09 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeifds7pfjepzufisnxjmntl3ewtn7jgsatw5hytv2ypzsj43apc6qm