OxCoder-9B-gptq-int4

OxCoder-9B — GPTQ Int4

Agentic coding, distilled to 9B, compressed to 8.5 GB.

OrionLLM/OxCoder-9B is a lightweight coding model for long-horizon agentic tasks — distilled from Fable-5.1 and GLM-5.3 trajectories across Claude Code, OpenCode and Codex — with 262K native context, SWE-bench Verified 73.5, Terminal-Bench 2.1 (Claude Code) 50.8. This artifact is a faithful GPTQ Int4 quantization: the full text stack at group size 128, desc_act ordering, calibrated on real code-reasoning and agentic tool-use traces — not generic web text.

Why this artifact

• Runs where the BF16 original can't. 17.6 GB BF16 → 8.5 GB Int4 fits a 12 GB GPU with room for KV cache. Same weights, same behavior, a third of the memory.
• Proven recipe, third shipped artifact on this hybrid qwen3_5 architecture (GatedDeltaNet + attention): 200/200 text modules quantized, mean per-layer quant error 2.3e-5 (worst 1.1e-4) — better than the 4.2e-5 baseline of the first artifact in the series. Vision tower skipped by design (text-only calibration).
• Calibrated on serving traffic. 256 rows: 128 OpenCodeReasoning CoT traces + 128 Nemotron-Agentic interactive tool-use episodes. The model sees, during calibration, the same kind of prompts it sees in production — code reasoning with think blocks, multi-step tool calls.
• Ships with the froggeric v22.5 chat template (froggeric/Qwen-Fixed-Chat-Templates): XML tool calls, preserved think blocks, KV-cache-safe history, no empty-think poisoning. Calibrated and served through the same template, so serving matches calibration exactly.

Serving (vLLM)

vllm serve malvavisc0/OxCoder-9B-gptq-int4 \
  --quantization gptq_marlin \
  --max-num-seqs 10 \
  --chat-template chat_template.jinja \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml

Notes:

• --max-num-seqs 10 — the hybrid GDN/Mamba cache corrupts state at high concurrency; 10 is the proven setting on this architecture.
• Pass --chat-template chat_template.jinja explicitly — vLLM does not auto-load the sibling file from the model dir in current releases.
• Sampling: temperature=0.6, top_p=0.95, top_k=20 (thinking mode). Greedy decoding causes repetition loops on this model class.

Hardware guidance

Quantization details

Toolchain: quantized with aft (the Aria Finetuner pipeline) driving gptqmodel 7.3.4 / torch 2.13.0+cu130 / transformers 5.15.0 on an NVIDIA GB10 (DGX Spark, aarch64, 128 GB unified). Quant error: mean 2.3e-5, worst 1.1e-4 (200 text modules).

Verified on consumer hardware: RTX A5000 16 GB (sm_86, plain CUDA — no GB10 dependency) with vLLM 0.28.0: 5/5 behavioral probes (math with closed think blocks, exec-verified Kadane's incl. the all-negative case, parsed XML tool call), 45 tok/s decode on a short coding probe, 40 tok/s at 31k context, 0.06 s TTFT.