GLM-5.3-Flash REAP-50 — GGUF

GGUF quantisations of zai-org/GLM-5.3-Flash with half of its routed experts removed by REAP saliency pruning (144 of 288 kept in every MoE layer). Five quant levels from 67.2 to 92.5 GiB, plus the vision projector, for people running the model in llama.cpp on a single 128 GB unified-memory machine. These files need the patched llama.cpp fork described below; stock llama.cpp cannot load them. The FP8 parent they were built from scores 0.8425 top-1 agreement with the unpruned model on 241,516 held-out tokens. The GGUF files themselves have not been evaluated that way.

Model overview

Which file should I use?

No measurement here ranks the quant levels by quality (see Evaluation), so choose by memory. Every file passed the same load / arithmetic / perplexity gate on a Jetson AGX Thor with -ngl 0 --no-repack.

Bits/weight are as reported by llama-quantize. Whether the imatrix was used is recorded in each file's own quantize.imatrix.* metadata.

Sibling repos (same v2 mask):

A later v3 mask (pass3-hope/) exists only inside those two repos. There is no v3 GGUF. v3 has never been evaluated end to end, and its healing used a fallback estimator (see Changelog).

Quickstart

1. Build the patched llama.cpp

These files declare general.architecture = glm5-next. The fork patrickbdevaney/llama.cpp adds that architecture: the mHC Sinkhorn operator, the hybrid KDA/MLA graph and the vision tower. All testing on this card used builds of that fork.

Upstream llama.cpp now also has a glm5-next architecture, but it reads different metadata keys. It expects glm5-next.hyper_connection.count and a required glm5-next.attention.indexer.kpool; these files carry glm5-next.attention.hc.mult and no kpool key. Upstream should therefore refuse to load them. That conclusion comes from reading upstream source (as of 2026-11-06), not from a test.

git clone --branch glm5-next-reap50-gguf-v1 https://github.com/patrickbdevaney/llama.cpp
cd llama.cpp
cmake -B build -DGGML_CUDA=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

The same changes ship in this repo in two other forms:

• glm5-next-llama.cpp.patch is 15 commits on upstream 761797ff; apply it with git am.
• glm5-next-llama.cpp.bundle is a self-contained git bundle with no prerequisite commits (git clone --branch glm5_next glm5-next-llama.cpp.bundle llama.cpp).

The tag sits at patch commit 12. Commits 13–15 (hidden-state dump, MTP draft and recurrent-state rollback) are on the glm5_next branch.

2. Run

Always add the two stop-token overrides. GLM-5.3 has three end-of-generation tokens (154820 <|endoftext|>, 154827 <|user|>, 154829 <|observation|>), but these headers carry only eos_token_id = 154820. Without the overrides, the model ends its answer with <|user|> and then writes a fabricated next turn. With them, llama.cpp lists all three as EOG tokens (verified in the vision runs).

Image + text, as run for the ChartQA check (Thor, CPU decode):

./build/bin/llama-mtmd-cli -m GLM-5.3-Flash-REAP50-IQ3_M.gguf \
    --mmproj mmproj-GLM-5.3-Flash-REAP50-F16.gguf --image chart.png -p "Your question" \
    -ngl 0 --no-repack -c 4096 \
    --override-kv tokenizer.ggml.eot_token_id=int:154827 \
    --override-kv tokenizer.ggml.eom_token_id=int:154829

Text, as run by the release gate (which ran without the overrides; add them for chat use):

./build/bin/llama-completion -m GLM-5.3-Flash-REAP50-Q4_K_M.gguf -c 512 -n 160 --temp 0 \
    -ngl 0 --no-warmup --no-repack -p "Q: What is 17 multiplied by 23?
A:"

llama-server with these files is untested.

On unified-memory machines (Jetson Thor and similar), start with -ngl 0 --no-repack. The file stays memory-mapped and its pages stay evictable. Offloaded layers become non-evictable device buffers, and -ngl 99 on a 93 GiB file ran the 128 GB Thor out of memory (recorded in scripts/gguf_gate.sh). Partial offload (-ngl 24, -ngl 36) of IQ3_M ran fine.

Evaluation

The parent: teacher-forced agreement with the unpruned model

Measured on the FP8 v2 checkpoint these GGUFs were converted from, not on the GGUFs. Both models are teacher-forced over 241,516 held-out tokens that calibration never saw (image-placeholder positions excluded). The metric is how often the pruned model's top-1 token matches the unpruned model's.

Top-k KL against the unpruned model is 0.6503. The vision bucket (532 text tokens) is too small to report. Source: artifacts/eval/eval_glm-5.3-flash-reap50-fp8-pass2.json in the build pipeline (glm-5.3-reap; method in wiki/97-evaluation.md). The damage is uneven: code and maths stay above 0.91, while general prose falls to 0.58.

Measured on these GGUF files

How to read these:

• The perplexity numbers detect a broken quant and nothing else. The gate text (artifacts/ppl_gate.txt) is printed PyTorch tensor reprs of token IDs, mostly digits, not natural language. The differences between levels are within noise and do not mean Q4_K_S beats Q4_K_M.
• ChartQA n = 24 is a smoke test. The confidence interval is wide, and the IQ3_M-vs-IQ4_XS gap is inside it. ChartQA years are graded exactly rather than within 5%.
• Sources: artifacts/ppl_GLM-5.3-Flash-REAP50-*.txt, artifacts/niah_GLM-5.3-Flash-REAP50-Q4_K_M.json, artifacts/vision_characterization.json (IQ3_M). The IQ4_XS ChartQA result is from the shipping log.

Not measured: teacher-forced agreement or KL for any GGUF file; the extra cost of 4-bit quantisation on top of pruning; generative benchmarks (MMLU, HumanEval, tool-calling, etc.); retrieval beyond 32,000 tokens; any quality ranking between the five levels.

Hardware and memory

Everything above ran on one NVIDIA Jetson AGX Thor (128 GB unified memory; llama.cpp reports 125,771 MiB of device memory) with fork builds of llama.cpp.

Long context is limited by compute, not memory. llama.cpp runs the 11 MLA layers as dense attention. The fork's DeepSeek-sparse-attention path exists but is off unless GLM5_DSA=1, and these files' imatrix has no indexer entries. NIAH latency grows from 54 s at 2k tokens to 236 s at 32k on Q4_K_M.

Speculative decoding with the built-in MTP block

Every file contains the model's native multi-token-prediction block as blk.45. make_mtp_draft.py (in this repo) extracts it into a standalone draft GGUF, which needs the glm5_next branch (not the tag) for recurrent-state rollback:

python make_mtp_draft.py GLM-5.3-Flash-REAP50-IQ3_M.gguf GLM-5.3-Flash-REAP50-IQ3_M-MTP-draft.gguf
./build/bin/llama-speculative-simple -m GLM-5.3-Flash-REAP50-IQ3_M.gguf \
    -md GLM-5.3-Flash-REAP50-IQ3_M-MTP-draft.gguf -ngl 24 --draft-max 3 --draft-min 1

Measured on IQ3_M at -ngl 24 with short (48-token) generations, it is slower than not speculating:

The draft itself costs 5.33 ms per token. The target does not get cheaper per token until the batch reaches 32 tokens (llama-bench: pp16 4.30 tok/s, pp32 23.05 tok/s), so verifying 2–4 tokens costs about as much as decoding them. Make the draft from the same file you serve: final hidden states differ between quant levels (cosine 0.891 between IQ3_M and IQ4_XS, against 0.999 for a same-quant control; research/MTP_CAPTURE_FINDINGS.md).

How it was made

1. Pruning (v2 mask). REAP saliency scores each routed expert by its router-weighted output norm over a calibration mixture. Calibration covered agentic, code, math, science, finance, general and image-text data. Licence policy: permissively licensed sources only. The 144 highest-scoring of 288 experts were kept in each MoE layer. Pruning deletes FP8 expert tensors and slices the router; surviving weights are not changed.
2. Healing. A measured per-layer scalar (42 layers; median 0.9102, range 0.8335–0.9495) was applied to the routed-expert block scales. A per-expert variant was tried first, measured worse, and was reverted on 2026-08-29, before these GGUFs were built.
3. Conversion. The FP8 checkpoint went to a Q8_0 GGUF (163.4 GiB, not published) through the fork's convert_hf_to_gguf.py with GLM5_KEEP_MTP=1, then to each level with llama-quantize --allow-requantize.
4. Importance matrix. imatrix-GLM-5.3-Flash-REAP50.dat (264,108,030 B; legacy format; 765 entries; 347 chunks of 512 tokens) was computed on the Q4_K_M file, not on Q8_0, and used for IQ4_XS, Q3_K_M and IQ3_M. Its calibration file (artifacts/imatrix_calib.txt) is, like the perplexity gate file, printed tensor reprs of token IDs rather than natural text. The activation statistics therefore come from digit strings. The effect on the three imatrix quants has not been measured. Some experts are never routed to by the calibration, so their rows fall back to unweighted quantisation (partial data (99.31%) in the imatrix log).

Files uploaded 2026-08-30 to 2026-08-31.

Limitations

• A patched llama.cpp is required (see Quickstart). Other GGUF runtimes have not been tried.
• General-prose and trivia knowledge took the most damage (top-1 agreement 0.58 on the parent). Code and maths held best.
• The quant levels are unranked. No teacher-forced, KL or benchmark comparison exists between them.
• The imatrix and perplexity-gate corpora are not natural text (see How it was made).
• The stop tokens are incomplete in the header; pass the two --override-kv flags.
• Sparse attention is not used by default, so attention is dense and long prompts are slow. Retrieval is verified only to 32k tokens.
• MTP speculation is a net slowdown in llama.cpp today.
• Vision evidence is limited to n = 24 ChartQA questions (IQ3_M and IQ4_XS only) and a synthetic-image smoke test on Q4_K_M (shapes, colours, positions and the number all identified correctly).

Changelog / corrections

• 2026-08-30 – 08-31. Q4_K_M, Q4_K_S, IQ4_XS, Q3_K_M, IQ3_M, mmproj, imatrix and llama.cpp patch uploaded.
• 2026-09-03. MTP draft support: make_mtp_draft.py, recurrent-state rollback, patch and bundle regenerated.
• 2026-10-12. Clarified that all GGUFs are v2 and no v3 GGUF exists.
• 2026-10-24. Corrections to the v3 description:
  • v3 was healed with the fallback first-moment per-layer scalar (median 0.766), not measured or per-expert gains.
  • v3 has not been evaluated end to end.
  • The v3 NVFP4 build also quantises the KDA attention projections.
  • None of this affects the GGUF files, which are v2.
• 2026-11-06. Card rewritten and verified against repo files and build artifacts; a 2026-10-26 automated rewrite contained errors.

License and attribution

MIT, the licence of the base model (LICENSE, Copyright (c) 2026 Z.AI Co., Ltd). The base model is GLM-5.3-Flash by Z.ai. Pruning follows REAP (Lasby et al., 2025, arXiv:2510.13999). Quantisation uses llama.cpp (MIT).

Citation

@misc{devaney2026glm53flashreap50gguf,
  author       = {Patrick Devaney},
  title        = {GLM-5.3-Flash REAP-50 GGUF: 50\% expert-pruned GLM-5.3-Flash for llama.cpp},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/patrickbdevaney/GLM-5.3-Flash-REAP50-GGUF}}
}