Qwen3 Coder Next
A large model in Alibaba's Qwen series targeted at code-related tasks, with a model file of approximately 52 GB. It is focused on code generation and code repair capabilities.
Timeline (last 90 days)
On the vLLM server side, you need to add the two parameters --enable-auto-tool-choice and --tool-call-parser qwen3_coder for Claude Code's tool calls to work properly.
Qwen3 Coder Next is an 80B-parameter mixture-of-experts (MoE) coding model, deployed by vLLM on a single A100 80GB, using the 4-bit AWQ quantized version published by cyankiwi.
This experiment was configured with a native context of 262,144 tokens, the KV cache stored in FP8, and a maximum of two concurrent sequences.
This experiment only ran smoke tests, did not measure the quality loss of the 4-bit quantized model, did not compare with the hosted version of Claude, and did not collect A100 throughput data.
Qwen3 Coder Next is an MoE model with 80 billion parameters and only 3 billion activated, scoring 70.2 on SWE-Bench Verified, but only 5.5 token/s in Atomic Chat testing.
All Facts (5)
On the vLLM server side, you need to add the two parameters --enable-auto-tool-choice and --tool-call-parser qwen3_coder for Claude Code's tool calls to work properly.
50%UnverifiedQwen3 Coder Next is an 80B-parameter mixture-of-experts (MoE) coding model, deployed by vLLM on a single A100 80GB, using the 4-bit AWQ quantized version published by cyankiwi.
50%UnverifiedThis experiment was configured with a native context of 262,144 tokens, the KV cache stored in FP8, and a maximum of two concurrent sequences.
50%UnverifiedThis experiment only ran smoke tests, did not measure the quality loss of the 4-bit quantized model, did not compare with the hosted version of Claude, and did not collect A100 throughput data.
50%UnverifiedQwen3 Coder Next is an MoE model with 80 billion parameters and only 3 billion activated, scoring 70.2 on SWE-Bench Verified, but only 5.5 token/s in Atomic Chat testing.
50%