Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

2.3bpw Qwen 27B checkpoint targets 200K context on 12GB Ampere cards

AI summary

Developer JakeATX released an updated 2.3bpw Qwen 27B checkpoint using the new AmperInteger+ quantization format, which combines fractional 2–3-bit weights, trellis-based encoding, and an RCO-like optimization. He reports that the checkpoint scored 84.3% on LiveCode Bench with a strict grader across three runs, compared with 83.9% for Qwen3.6 27B BF16 and 90.3% for Qwen3.8 BF16. Using a custom SJKVaRN build and runtime changes, he says the 7.4GiB on-card checkpoint can support 200K-token context on a single 12GB Ampere card.

In-depth insight

The release combines a concrete quantized model artifact, a new quantization format, and a claimed 200K-context configuration for 12GB Ampere hardware, making it relevant to local model deployment.

#JakeATX#LlamAmpere#AmperInteger+#Qwen3.8-27B-Swift1.5-Uncensored-ATX-AI2#Qwen3.6#Qwen3.8#SJKVaRN#LiveCode Bench#Swift-1.5#Mirai

Original · 2 min readRead the original
Group chat

No comments yet. Start the conversation.