Why this model exists
The full-precision GLM-5.2 de-risked master is ~1.35 TiB and needs an 8×B200-class node. The NVFP4
build brings that to ~451 GB. This build goes further — a non-uniform REAP prune on top of
NVFP4, landing at ~323 GB, roughly 28% smaller than its NVFP4 parent and about a quarter
of the BF16 master.
The point is footprint. If the NVFP4 build is close but doesn't fit your envelope, this is the one
to ask about.
Specifications
Table | |
|---|
| Architecture | GlmMoeDsaForCausalLM — GLM MoE with MLA + DeepSeek-style Sparse Attention |
| Base | zai-org/GLM-5.2 — official |
| Parent | GLM-5.2-DERISKED-NVFP4 |
| Transform | Refusal-direction de-risk (in parent) → NVFP4 → non-uniform REAP prune |
| Keep set | NU176 — non-uniform, per-layer budgets rather than a flat keep |
| Precision | NVFP4 (4-bit floating-point, microscaled, two-level block scaling) |
| On-disk | ~323 GB · 65 safetensors shards |
| Languages | English, Chinese |
| Serving | vLLM |
| Status | EXPERIMENTAL — unbenchmarked |
What "NU176" means
REAP is router-weighted expert activation pruning: experts are scored by their contribution
across a calibration set, and the least-salient are dropped.
NU denotes a non-uniform keep — the per-layer budget varies with measured routing load rather
than removing the same count from every MoE layer. Layers that genuinely spread work across many
experts keep more of them; layers whose routing concentrates keep fewer. That buys a better
capability-per-byte trade than a flat prune at the same total size, at the cost of less predictable
per-layer memory.
Contrast with the Kimi K3 line, where
KIMI-K3-DERISKED-CODER-REAP-320-MXFP4
uses a uniform 320-of-896 keep specifically so memory and routing stay predictable layer to layer.
Both approaches are available; which is right depends on whether you are optimising for quality at a
size target or for predictability.
Lineage
zai-org/GLM-5.2 official open-weight, 256 routed experts
└─ GLM-5.2-DERISKED-BF16 in-house refusal-direction de-risk, from source
└─ GLM-5.2-DERISKED-NVFP4 NVFP4 quantization
└─ GLM-5.2-DERISKED-REAP-NU176-NVFP4 ← this repo, non-uniform REAP prune
Table | |
|---|
| Applied | Weight-level de-risk (in BF16 master) · NVFP4 quantization · non-uniform REAP prune |
| Not applied | SFT · DPO · RLHF |
| Format | HF safetensors · NVFP4 |
On refusal behaviour: this checkpoint inherits the parent's deliberately reduced refusal
surface. It is a Blackfrost de-risked model. Do not evaluate or rate-limit it as if it were a
safety-stock derivative of upstream GLM-5.2.
Measured behaviour
Pending
Table with columns: Benchmark, GLM-5.2-DERISKED-NVFP4, This (REAP-NU176), Retention| Benchmark | GLM-5.2-DERISKED-NVFP4 | This (REAP-NU176) | Retention |
|---|
| pending | — | — | —% |
Expert pruning is a capability trade, and a card that ships a pruned build without publishing what
it cost is asking the reader to take the trade on faith. Harness, conditions and retention figures
will be stated here — including any benchmark where retention is poor.
Two things specifically unmeasured on this checkpoint:
- Capability retention versus the NVFP4 parent — the prune's actual price.
- Whether the de-risking survives the prune. The parent's refusal figures were measured before
pruning. Pruning removes experts; whether the refusal-direction intervention persists unchanged
through that is an open question on this artifact, not an established property.
Until both are measured, the parent's numbers should not be quoted for this build.
Deployment notes
- Hardware. ~323 GB of weights. Fits comfortably on a single 8×96 GB node with room for KV, and is the only build in this line that will.
- Runtime. vLLM. NVFP4 kernels require sm_120-class hardware (RTX PRO 6000 Blackwell) or newer; sm_100 (B200) does not carry NVFP4 MoE kernels — use the MXFP4 or BF16 builds there.
- Thinking. GLM-5.2 reasoning output may land in the visible stream depending on chat-template handling. Budget
max_tokens generously.
- Integrity. Verify shard count (65) and byte totals after download before attributing a load failure to the weights.
Other builds
Blackfrost publishes the GLM-5.2 de-risked line in several shapes. Access to full-precision and
custom builds is granted on purchase under a separate licence agreement — those repositories
are gated, and approval follows a completed agreement. The gate is the transaction.
Table with columns: Build, What it is| Build | What it is |
|---|
GLM-5.2-DERISKED-BF16 | Full-precision BF16 master — every other build in this line derives from it |
GLM-5.2-DERISKED-NVFP4 | 4-bit NVFP4, single-node serving on sm_120-class hardware |
| This repo | Non-uniform REAP prune on top of NVFP4 — smallest footprint in the line |
| Custom | Expert budgets, uniform or non-uniform keep sets, or de-risking calibrated against your own threat model |
Purchase link coming soon. Until then, @Blackfrost_AI DMs are
the fastest route to a human.
Disclaimer
Refusal behaviour in this checkpoint has been deliberately modified at the weight level. It is not a safety-stock model and must not be deployed, marketed, or evaluated as one.
Capability has been deliberately reduced. Experts were removed by non-uniform prune. The size of that reduction has not been measured on this checkpoint. Do not assume parent-equivalent quality.
No warranty of any kind. Provided "as is", without warranty express or implied, including fitness for a particular purpose. Nothing here guarantees that any given input will be accepted or refused, that any capability is retained, or that any category of output is unreachable.
No measurements are published for this build. Figures on the parent card describe the parent, under its own stated harness and caveats, and do not transfer here.
Modification by a recipient voids this characterization. Blackfrost's obligations attach at the point of release. Any further ablation, fine-tuning, merging, quantization or alteration by a recipient produces an artifact Blackfrost has not evaluated and does not stand behind — responsibility for that artifact transfers entirely to whoever produced it.
Operator-owned policy. Open weights mean the operator sets and enforces policy. Deploy only in controlled environments with access control, independent logging and review.
Access & licensing
Access is granted automatically on purchase — you do not wait on a manual review.
➜ Purchase access to this model — enter your Hugging Face username at checkout, and your account is granted access to this repository within moments of payment.
- Base licence: MIT, inherited from
zai-org/GLM-5.2 — upstream terms travel with this derivative.
- Blackfrost terms: commercial licence, layered on top of the upstream licence.
- Redistribution: do not redistribute weights outside your grant.
- Evaluation recommendation: should not be evaluated by processes that assume refusal behaviour equivalent to upstream GLM-5.2.