kat-coder-v25-dev-pqm

Standalone Kwaipilot/KAT-Coder-V2.5-Dev (clean upstream code / tool-calling model, Qwen3.6-35B-A3B MoE) exported to .pqm for the prism-engine inference server. GGUF-free standalone boot — this single .pqm container is self-contained and boots directly on prism-engine, with no GGUF, safetensors, or transformers checkpoint alongside it.

License & attribution

This repository distributes a composite work under a dual arrangement:

• Proprietary layer — © 2026 BCZ Singapore Pte Ltd. All rights reserved. The .pqm container format and packaging, and the prism-engine inference server (Rust/CUDA) are proprietary and are not licensed under Apache-2.0.

• Base model weights — Apache-2.0. The underlying weights are an unmodified (repacked-only) derivative of a single Apache-2.0 upstream:

  Component	Source	Copyright	License	Modification
  Full model (backbone + experts)	Kwaipilot/KAT-Coder-V2.5-Dev	© 2026 Kwaipilot / Kuaishou	Apache-2.0	weights repacked to .pqm (mixed Q4_K/Q6_K/F32); no fine-tune, no merge

  KAT-Coder-V2.5-Dev (the open Apache-2.0 weights, itself Qwen3.6-35B-A3B-based) — not the separate closed API-only "KAT-Coder" product. These base weights remain licensed under Apache-2.0. The full Apache-2.0 license text and per-component attribution are in the NOTICE file in this repo.

This is not an original, trained-from-scratch foundation model — it is a quantized repack of the Apache-2.0 upstream named above. Use of the base weights is governed by Apache-2.0; use of the proprietary .pqm packaging and prism-engine requires a separate license from BCZ Singapore Pte Ltd.

Requirements — read before downloading

• Inference engine: .pqm is a proprietary container, not readable by vLLM, llama.cpp, Ollama, or transformers. Runs only on prism-engine (proprietary Rust/CUDA server, not included in this repo). Contact sprappcom for engine access.
• Tokenizer sidecar: .pqm does not embed a tokenizer. Use the kat_v25.tok shipped in this repo via PRISMX_TOKENIZER.
• Hardware: A100 (sm_80) / Ada-Ampere (sm_86/sm_89) validated. Full residency ~20 GB VRAM, or run with CPU-RAM expert offload on smaller cards.

Architecture

• Qwen3.6-35B-A3B MoE, hidden_size=2048, 40 layers (30 Gated DeltaNet + 10 full attention)
• 256 routed experts + 1 shared expert, top-8 routing
• vocab_size=248320

Quantization

Tensor group	Type
token embedding, attn_qkv/attn_output/attn_gate, routed experts (ffn_*_exps)	Q4_K
shared expert (ffn_*_shexp), output.weight	Q6_K
norms, SSM gates	F32

~4.56 BPW. Container ~39.9 GB (routed experts stored raw for CPU-offload support).

Usage

export PRISMX_PQM_STANDALONE=1
export PRISMX_PQM=/path/to/kat_v25.pqm
export PRISMX_TOKENIZER=/path/to/kat_v25.tok
prism_server 0.0.0.0:8080 \
  --n-cpu-moe 28 \
  --max-prefill-chunk 8 \
  --moe-cache-experts 2048 \
  --max-batch 1

Standalone mode always binds 0.0.0.0:8080. OpenAI-compatible /v1/completions and /v1/chat/completions. Use temperature >= 0.15 (greedy decoding can collapse).

Verification

Boot-verified GGUF-free on A100 (sm_80) with prism_server (PRISMX_PQM_STANDALONE=1 --n-cpu-moe 28 --max-prefill-chunk 8 --moe-cache-experts 2048 --max-batch 1, tokenizer kat_v25.tok from this repo): coherent English code and tool-calling completions, no NaN/garbage output.

Known limitations

• Text-only. This .pqm carries no vision weights; image input is available on the separate AWQ variant, not this file.
• Intermittent sub-word token drop/duplication on some code/arithmetic prompts is a known open issue; validate structured output downstream.

Integrity

File	Bytes	SHA-256
kat_v25.pqm	39870623744	0050a596155bee84d162148a3e9248099f2dc838c41dca1391eb5a33ef693b15
kat_v25.tok	8949400	md5 907b88ec3a1f5aed0e234472eb441520

Check: sha256sum kat_v25.pqm must print the hash above.