Ttimms/KAT-Coder-V2.5-Dev-REAP-50-GGUF

KAT-Coder-V2.5-Dev · REAP-50 — GGUF

GGUF quants of the 50%-REAP-pruned KAT-Coder-V2.5-Dev (Gated-DeltaNet + MoE hybrid): 256 → 128 experts, ~19B parameters, no MTP head.

• Flagship (NVFP4A16, vLLM): Ttimms/KAT-Coder-V2.5-Dev-REAP-50-NVFP4A16
• Pruned bf16 source: Ttimms/KAT-Coder-V2.5-Dev-REAP-50-bf16
• Pipeline + benchmarks: https://github.com/t-timms/kat-coder-nvfp4

Base-model quality (NVFP4A16 quant, greedy, instruct): HumanEval+ ~90%, MBPP+ ~90%. These are single greedy draws; the accuracy harness has a measured run-to-run spread of 1.85–4.27 pp per task (2026-09-05), so read small differences as unresolved.

Architecture

graph TD
    Base["Kwaipilot/KAT-Coder-V2.5-Dev<br/>Qwen3.5 MoE - 256 experts - ~69 GB bf16"]
    subgraph Build ["Build - RTX 5070 Ti, SM120"]
        REAP["REAP expert prune 50%  256 -> 128 experts + router-renorm fix"]
        Strip["strip vision tower + 333 untrained tensors"]
        Quant["NVFP4A16 quantize (weight-only, data-free, 82 s)"]
    end
    subgraph HF ["Published formats"]
        A16["REAP-50-NVFP4A16 - 12.45 GiB (default, vLLM)"]
        W4A4["REAP-50-NVFP4-W4A4 (native FP4)"]
        GPTQ["REAP-50-NVFP4A16-GPTQ (null result, kept for transparency)"]
        GGUF["REAP-50-GGUF (Q4_K_M / Q5_K_M / Q6_K / Q8_0)"]
        BF16["REAP-50-bf16 (pruned source)"]
    end
    Bench["A16 single draws - HumanEval+ 89.0% - MBPP+ 90.5% - SWE-bench Verified 52.0%"]
    Base --> REAP --> Strip --> Quant --> A16
    Strip --> BF16
    BF16 -. re-quant .-> W4A4
    BF16 -. re-quant .-> GPTQ
    BF16 -. convert .-> GGUF
    A16 --> Bench

Download

Grab one file, not the whole repo.

File	Quant	Size	Notes
kat-reap50-Q4_K_M.gguf	Q4_K_M	11.4 GB	recommended — fits 16 GB with room for context
kat-reap50-Q5_K_M.gguf	Q5_K_M	13.3 GB	higher quality, tighter KV budget
kat-reap50-Q6_K.gguf	Q6_K	15.3 GB	near-lossless; little headroom on 16 GB
kat-reap50-Q8_0.gguf	Q8_0	19.7 GB	needs >16 GB or CPU offload

MoE with ~3B active parameters — CPU / Apple-Silicon inference is fast for the size. For full precision, re-quantizing, or merging, use the pruned bf16 source and run convert_hf_to_gguf.py --no-mtp yourself.

Run it in

• llama.cpp — llama-server -m <file>.gguf -ngl 99
• LM Studio (bundles a recent llama.cpp)
• koboldcpp
• Jan
• text-generation-webui
• Ollama — once its bundled llama.cpp includes this arch (see Requirements)

Measured on an RTX 5070 Ti (16 GB)

Measured 2026-09-24 on kat-reap50-Q4_K_M.gguf only (the other quants were not run). Plain decoding: this model has no MTP head, so there is no speculative decoding. Hardware: RTX 5070 Ti (16,303 MiB), Ryzen 5 7600, WSL2, with the desktop still using 1.3-1.9 GB of VRAM. Software: llama.cpp commit 3173a56 (source build, 2026-08-29).

llama-server --model kat-reap50-Q4_K_M.gguf -c 32768 -ngl 999 -fa on -np 1 \
  --cache-type-k q8_0 --cache-type-v q8_0 -b 2048 -ub 512 --jinja

Metric	Result
Decode, empty context (T = 0 / 0.6 / 1.0)	181.0 / 183.6 / 183.5 tok/s
Decode at ~24K tokens of context (T = 0.6)	151.9 tok/s
Prefill, ~24K-token prompt	~4.7K tok/s (about 5 s)
Peak VRAM at 32K context, q8_0 KV	12,873 MiB of 16,303 MiB (includes the desktop's 1.3-1.9 GB)

Decode figures are medians of the server-reported predicted_per_second: 5 prompt types x 3 repeats at empty context, 2 prompt types x 2 repeats at ~24K context. One machine and one build, so treat these as a reference point rather than a guarantee.

Requirements

Needs a recent llama.cpp (the qwen35moe hybrid arch). Verified with llama-server from a source build. Ollama support pending its bundled-llama.cpp bump.

llama-server -m kat-reap50-Q4_K_M.gguf -ngl 99 --port 8080

The chat template defaults to thinking on; pass "chat_template_kwargs": {"enable_thinking": false} for direct output.

License

Apache-2.0 (inherits from Kwaipilot/KAT-Coder-V2.5-Dev). Pruning: REAP (github.com/CerebrasResearch/reap, with a router-renormalization fix).