GLM-4.7-Flash-24B-A3B-KREAP

GLM-4.7-Flash-24B-A3B-KREAP — Force-Fit Ablation

Korean-priority one-shot MoE Expert pruning of zai-org/GLM-4.7-Flash, produced with K-REAP.

한국어 요약

Routed Experts were pruned 64 → 48 per layer (25.0%) while top-4 routing was kept unchanged. This percentage is the reduction in routed Expert count, not total model parameters. A3B follows the base model's approximate active-parameter class because top-4 routing is unchanged, not a new active-parameter measurement The checkpoint contains 24.27B parameters (base 31.22B; 22.25% total-parameter reduction) and occupies 45.23 GiB (approximately 22.2% smaller than the base checkpoint), is stored in BF16/F32 safetensors, and received no fine-tuning.

Force-fit KREAP research checkpoint, not a strict hard-protection result — one layer exceeded the protected-set budget and six protected Experts were force-dropped.

Intended use

• Research on Korean-aware MoE Expert pruning and retention–compression trade-offs
• Comparative evaluation against the base model and matched-size pruning baselines
• Offline experimentation with the validated Transformers and PyTorch/CUDA path on sufficiently provisioned hardware
• Not intended for production deployment, safety-critical use, or claims beyond the reported evaluations

What this checkpoint shows

• Method: KREAP force-fit ablation
• Korean MC macro: 62.8 (base 68.8, -6.0 pp)
• Overall non-Korean composite: 56.4 (base 65.5, -9.1 pp)
• Coding 2-benchmark macro: 70.9 (base 76.0, -5.1 pp)
• LiveCodeBench v6 pass@1: 20.6 (base 22.9, -2.3 pp)
• Coding 3-benchmark macro: 54.1 (base 58.3, -4.1 pp)
• Protection budget: One layer exceeded the 48-Expert budget; 6 protected Experts were force-dropped.

Benchmark results

Aggregate rows are unweighted arithmetic means. The overall non-Korean composite averages six English reasoning/knowledge metrics, two math metrics, and two safety/bias metrics. The coding 3-benchmark macro averages HumanEval+, MBPP+, and LiveCodeBench v6 pass@1. These mixed summaries are compact internal comparisons, not standardized leaderboard scores.

Compatibility

• Architecture: Glm4MoeLiteForCausalLM.
• Model code: native Transformers implementation; no repository-side modeling_*.py is required, and the checkpoint loads with trust_remote_code=False.
• Preserved: first Dense layer, shared Experts, topk_method=noaux_tc, norm_topk_prob=true, router correction bias, and routed_scaling_factor=1.8.
• Routing: top-4 remains unchanged; only routed Expert tensors and corresponding router rows/bias entries were compacted with one old→new ID map per layer.
• The extra layer-47 MTP/NextN tensors are retained for checkpoint completeness, but runtime validation reports mtp_compatible=false. Use standard autoregressive generation; speculative MTP/NextN decoding is not supported by this release.
• Top-4 routing is unchanged, so the checkpoint-size and resident-weight reduction must not be interpreted as a proportional reduction in active Expert compute or latency. Cold-load time, TTFT, prefill/decode throughput, batch scaling, power, and 4K/32K/128K serving behavior have not been benchmarked.

How it was made

Source checkpoint

• Base: zai-org/GLM-4.7-Flash
• Frozen revision: 7dd20894a642a0aa287e9827cb1a1f7f91386b67
• Original routed MoE layout: first layer Dense, then 46 routed MoE layers with 64 Experts, top-4 routing, plus shared Experts.

Calibration dataset

• 6,848 records contain source teacher responses.
• 1,344 GLM-specialized records use deterministic base-model rollouts.
• Prompt/response observations cover 3,898,756 token memberships.
• The manifest passed a separate internal token-length and exact-prompt contamination validator.
• The exact calibration manifest and deterministic reconstruction script are not included in this model repository, so the full GLM experiment is not yet externally reproducible from public materials alone.

Observation and scoring

• Selected-only observer: only routed Expert assignments were measured; no dense [experts, tokens, hidden] activation tensor was retained.
• REAP contribution: normalized dispatch weight × L2 norm of the unweighted Expert output.
• Router selection and dispatch were checked separately so GLM's correction bias affects selection without being mistaken for the actual dispatch weight.
• The observed model remained logit-identical to the uninstrumented eager model in validation.

Pruning

• Uniform routed Experts per layer: 64 → 48.
• Shared Experts and the first Dense layer were not pruned.
• Top-4 and routed scaling factor 1.8 were retained.
• Safetensors surgery physically compacted Expert tensors, router rows, and correction-bias entries.
• Protection result: One layer exceeded the 48-Expert budget; 6 protected Experts were force-dropped.

Evaluation scope

These are internal comparative evaluation results, not standardized leaderboard numbers:

• KMMLU: deterministic 500-example subset (seed 42).
• KoBEST: declared BoolQ, COPA, SentiNeg, and HellaSwag evaluation splits.
• ARC-Challenge, HellaSwag-en, WinoGrande, MMLU-Pro, TruthfulQA MC2, MMLU-en, GSM8K, MathQA, BBQ, and SafetyBench: deterministic 512-example subsets.
• HumanEval+: full 164 tasks; MBPP+: frozen full 378-task revision; greedy pass@1.
• LiveCodeBench: full incremental v6 split (175 problems), n=1 greedy no-thinking generation, batch size 1, with official functional tests executed in an isolated ARM64 container. This protocol is not directly comparable to the official leaderboard's n=10 sampling protocol.
• The machine-readable field decision_eligible=true means that the run passed the project's internal screening checks. It is not an external certification. release_eligible=false.
• Do not compare these subset scores directly with full-set or few-shot leaderboard values.

Machine-readable results and the exact survivor manifest are included under evaluation/.

Internal release-gate result: The project plan's 25% gate allows at most -1.5 pp Korean macro and -2.5 pp overall non-Korean composite. This checkpoint measured -6.0 pp and -9.1 pp and force-dropped six protected Experts, so it did not pass the internal release criteria.

Interpretation boundaries

• At the same 48-Expert size, this checkpoint's Korean MC macro is 9.3 pp above KREAP-Code-48E, while its overall non-Korean composite is 4.0 pp lower, LiveCodeBench is 4.6 pp lower, and its coding 3-benchmark macro is 3.6 pp lower. This is a trade-off, not evidence that one checkpoint is universally better.
• This is not a controlled single-variable ablation: KREAP-48E and KREAP-Code-48E differ in both calibration data and survivor-selection rules. A same-calibration REAP baseline and component ablations are still required to isolate hard protection.
• Machine-readable results include direct KREAP-48E versus KREAP-Code-48E McNemar tests for 11 non-Korean binary/pass-fail tasks. Direct paired tests for the Korean benchmark set, aggregate bootstrap confidence intervals, multiple-comparison correction, and repeated calibration seeds are absent.
• PASS in the artifacts refers to checkpoint integrity and internal pipeline validation, not external replication.

Usage

Install a Transformers release that includes Glm4MoeLiteForCausalLM:

pip install "transformers>=5.0.0" accelerate safetensors

The checkpoint was validated with Transformers 5.0.0 and NVIDIA PyTorch 25.09. It uses the native Transformers model implementation and does not execute custom code from this repository.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "RAS-LAB/GLM-4.7-Flash-24B-A3B-KREAP"
tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=False,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=False,
    attn_implementation="eager",
    experts_implementation="eager",
)
model.eval()

messages = [{"role": "user", "content": "한국어로 이진 탐색을 설명해줘."}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

print(tokenizer.decode(output[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Limitations

Expert pruning is irreversible within this checkpoint and can change routing behavior sharply on tasks not represented by the calibration or evaluation sets. No post-pruning recovery or fine-tuning was applied. Reported Korean behavior is relative to the tested base checkpoint and harness; it does not imply universal Korean-language safety or quality preservation. The current release is not validated for speculative MTP/NextN decoding.

The Korean evaluation is primarily multiple-choice. Korean long-form generation, summarization, multi-turn dialogue, format-following, code explanation/debugging, factuality, hallucination, safety, tool calling, JSON output, and long-context retrieval were not comprehensively evaluated. Chinese capability is inherited from the base model but was not specifically protected or evaluated after pruning. vLLM/SGLang compatibility and deployment performance remain unverified.

Included research artifacts

• evaluation/glm-full-b20-results-with-livecodebench.json: machine-readable five-model full results

• evaluation/glm-full-b20-long-scores.csv: compact long-format B20 scores

• evaluation/glm-full-b20-independent-validation.json: independent B20 validation report

• evaluation/survivor-manifest.json: exact per-layer old Expert IDs retained

• artifacts/calibration-validation.json: separate internal calibration-manifest validation

• artifacts/observation-validation.json: selected-only observation validation

• artifacts/expert-statistics.csv: per-Expert scoring/features used for this method

• artifacts/protected-experts.json: exact KREAP protected set and evidence

Provenance

• Base model: zai-org/GLM-4.7-Flash, revision 7dd20894a642a0aa287e9827cb1a1f7f91386b67
• Base license: MIT
• General method code: K-REAP, commit 234666e
• Reproducibility note: the linked repository documents K-REAP and its public adapters, but the exact GLM-specific observation pipeline, run configuration, and evaluation manifests used for this checkpoint are not yet published there.
• Team: RAS-LAB
• Benchmark report SHA256: 71ab2ce742bf8a481d90d10c8b8684a918241fd475729b0bf2afe56b44335dd6
• Model card generated from an internally validated benchmark artifact with zero validation failures; this is not an external reproduction claim.

Citation

Please cite the base checkpoint, REAP, and K-REAP when using this model:

@misc{zai2026glm47flash,
  title        = {GLM-4.7-Flash},
  author       = {{Z.ai}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/zai-org/GLM-4.7-Flash}},
  note         = {Base revision 7dd20894a642a0aa287e9827cb1a1f7f91386b67}
}

@article{lasby2025reap,
  title   = {REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author  = {Lasby, Mike and Lazarevich, Ivan and Sinnadurai, Nish and Lie, Sean and Ioannou, Yani and Thangarasa, Vithursan},
  journal = {arXiv preprint arXiv:2510.13999},
  year    = {2025},
  doi     = {10.48550/arXiv.2510.13999}
}

@software{chedrian07_2026_kreap,
  author  = {Chedrian07},
  title   = {K-REAP: Language-preserving one-shot MoE expert pruning},
  year    = {2026},
  url     = {https://github.com/Chedrian07/K-REAP},
  version = {234666e0dde8714b0c61f36b67f6f9f053c02d07}
}