AI

Basalt 0.4.0: More models, still as fast - Qwen 27B up to 770 tok/s on a 5090

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 15/11/2026, 22:47:26 UTC.

Briefings

Basalt 0.4.0 released November 15, 2026, introducing the first dense Qwen 3.8 27B model and multi-model support. The model achieves up to 776 tokens per second on structured output and 262 on conversation, matching previous 246 tokens per second on structured output. It features video and file support with a 262K context window, maintaining 92.8% top-1 accuracy against BF16. The update includes router support for DeepSeek V4.1 Flash and various bug fixes.

Key points

• Basalt 0.4.0 released November 15, 2026, introducing the first dense Qwen 3.8 27B model with up to 776 tokens per second on structured output.
• The model supports video and file inputs with a 262K context window, matching the 246 tokens per second on structured output achieved with DeepSeek V4.1 Flash.

Context

This update represents a significant step in Basalt's development, transitioning from a router-based architecture to a dense model architecture while maintaining high performance metrics. The release coincides with the preparation for Qwen 4, allowing for early benchmarking of dense kernels.

Limitations

Performance metrics are snapshot-based and may vary with context length; benchmark data for video and file support is not fully detailed in the source.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 15/11/2026, 22:48:22 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeihltimhjxye4nd5rzvlmzsy23dd23sjbu6u442o2nfybku6g5e6te