GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
drowzeys
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
drowzeys
Model Tree
RedHatAI/GLM-5.3-Flash-NVFP4
Input Modalities
Output Modalities
Supported Functionality
Keys learned this on DeepSeek-V4-Flash 0731. Projecting residual writes through early layers made the target stop refusing and made the stock drafter keep proposing refusal-shaped tokens.
Spare early layers. That is still our default. On GLM-5.3-Flash the in-checkpoint drafter is MTP layers.45, and residual refusal lived late + MTP. Rank-1 projection of Blackfrost’s V (nearly orthogonal to Dealign, cos ≈ 0.09) never cleared 9/32, even widened to L2–45.
The previous publish of this same repo id was a LibertAI ModelOpt 120-shard body with Dealign o_proj including L44. Two changes in this drop:
| Blackfrost DERISKED-NVFP4 (reference) | Dealign / previous keys (LibertAI body) | OrcaRouter Uncensored NVFP4 (guide) | This checkpoint | |
|---|---|---|---|---|
| Stock quant | Their NVFP4 of derisked BF16 | LibertAI ModelOpt 120-shard | compressed-tensors, no MTP | RedHat compressed-tensors |
| Edited layers | skip_early=2 → L2–44 o_proj | L15–45 o_proj (incl. L44) | full uncensored swap | L15–43 + L45 o_proj (30 tensors) |
| L44 | edited | Dealign Δrel 0.74 | n/a | RedHat stock (anchor-safe) |
| MTP L45 |
Abliteration does not change FLOPs. The intended win is direct completions instead of refuse/hedge loops, without soup.
RedHatAI/GLM-5.3-Flash-NVFP4 is the parent: LLM-Compressor compressed-tensors NVFP4 on routed-expert FFNs, attention / vision / embeddings BF16, ~193 GiB, 10 shards + MTP. Experts, vision, QKV, embeddings, L0–14 o_proj, and L44 o_proj remain theirs.
orcarouter/GLM-5.3-Flash-Uncensored-NVFP4 — compressed-tensors experts-only NVFP4, attention BF16. We used it as a guide for the quant split. We did not ship their checkpoint (they drop MTP).
Full credit to dealignai / @dealignai (compute @jordanschenck) for GLM-5.3-Flash-UNCENSORED-NVFP4.
We byte-copied BF16 self_attn.o_proj for L15–43 and MTP L45 onto RedHat shards. L44 is not copied. We did not ship their full checkpoint as a swap.
Full credit to Blackfrost / @Blackfrost_AI for publishing GLM-5.3-Flash-DERISKED-NVFP4 and the BF16 master it comes from.
We recovered their rank-1 residual V by SVD of (Blackfrost − LibertAI) o_proj. It is nearly orthogonal to Dealign’s write (cos ≈ 0.09). Projecting it on L10–35, L2–35, or L2–45 never passed 9/32. Credited as a measured reference. Do not treat this repo as a Blackfrost build.
Glm5NextForConditionalGeneration, hybrid KDA+DSA, mHC, native MTP).--block-size 2304; barrydeen GMU 0.85 floor.Byte-copy Dealign o_proj into RedHat shards. Not a λ-projection of a rank-1 V. Offsets from RedHat headers — never the LibertAI 120-shard map.
| Tensor | model.language_model.layers.{L}.self_attn.o_proj.weight (BF16) |
| Layers | 15–43 and 45 (30 tensors, includes MTP layers.45) |
| Skip | L44 RedHat stock (Dealign L44 Δrel 0.74) |
| Safety | L0–14 byte-identical to RedHat stock |
| Experts | NVFP4 passthrough (even in rewritten shards) |
| Gate | 32/32 bypass, 0 refuse, 0 garble, raw vLLM + thinking-off template |
Variation table: GitHub METHOD.md. Artifacts: ABLIT_META.json, VARIATIONS.json.
Reproduce:
bash
python3 scripts/apply_oproj_l15_45.py \--src /path/to/GLM-5.3-Flash-NVFP4-RedHat \--dst /path/to/dest \--bins ./oproj_bins \--skip-layers 44 \--fresh
Stock GLM-5.3-Flash always opens <think> and injects Reasoning Effort: Max. enable_thinking=false is a silent no-op — CoT lands in content. Mount chat_template.thinking-off.jinja over chat_template.jinja at serve time. Stock chat_template.jinja in this repo is unchanged from RedHat.
| Path | Purpose |
|---|---|
model-00001-of-00010.safetensors … 00010 + model_mtp.safetensors + index | Full NVFP4 checkpoint (RedHat layout, L15–43+L45 o_proj from Dealign) |
ABLIT_META.json | Edit stats / recipe fingerprint |
VARIATIONS.json | Refusal32 table |
chat_template.thinking-off.jinja | Recommended serve overlay (closed <think></think>) |
config.json is RedHat stock (index_topk 2048, quant_method=compressed-tensors). Honor it — do not pass --quantization modelopt_fp4.
The previous 120-shard LibertAI ModelOpt files (model-*-of-00120.safetensors) are removed from this repo.
bash
# after you agree to the gate (automatic approval)hf download drowzeys/keys-GLM-5.3-Flash-NVFP4-ablit-l15-43-mtp-l45 \--local-dir ~/models/GLM-5.3-Flash-NVFP4-ablit-l15-43-mtp-l45
Serve with Tony’s 2× DGX Spark DFlash2 recipe (marlin MoE, DFlash2 k=7, fp8 KV, --block-size 2304). GPU memory utilization ≤ 0.85. Mount the thinking-off template.
MIT, inherited from Z.AI GLM-5.3-Flash (also the RedHat NVFP4 card). You must still comply with the Responsible Use gate above.
| BF16 exact passthrough (~10 GiB extra) |
Dealign o_proj |
| dropped |
RedHat NVFP4 experts, Dealign o_proj |
| Refusal32 | not our gate | 32/32 (previous publish) | their card | 32/32 bypass, 0 refuse, 0 garble |
tokenizer.json / chat_template.jinja / processor | Unchanged from RedHat / Z.ai |