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
LibertAIDAI/GLM-5.3-Flash-NVFP4
Input Modalities
Output Modalities
Supported Functionality
Keys learned this on DeepSeek-V4-Flash 0731. DSpark taps L40–42 live in that checkpoint. Projecting residual writes through L10–42 made the target stop refusing and made the stock drafter keep proposing refusal-shaped tokens — accept died, tok/s left the stock band.
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. Dealign’s card is explicit: if MTP o_proj stays stock, the draft head keeps proposing refusals. Rank-1 projection of Blackfrost’s V (nearly orthogonal to Dealign, cos ≈ 0.09) never cleared 9/32, even widened to L2–45.
So this drop:
o_proj including MTP — required for 32/32 on this model.| Blackfrost DERISKED-NVFP4 (reference) | Dealign UNCENSORED-NVFP4 (o_proj donor) | This checkpoint | |
|---|---|---|---|
| Stock quant | Their NVFP4 of derisked BF16 | Same ModelOpt layout as LibertAI | LibertAI body + Dealign o_proj L15–45 |
| Edited layers | skip_early=2 → L2–44 o_proj | L0–11 stock, late + MTP o_proj | L15–45 o_proj (31 tensors) |
| MTP L45 | BF16 exact passthrough (~10 GiB extra) | NVFP4 experts, o_proj edited | LibertAI NVFP4 experts, Dealign o_proj |
| Refusal32 | not our gate | HarmBench 320/320 (their card) | 32/32 bypass, 0 refuse |
| Goal on GB10 | Enterprise derisked artifact | Uncensored LibertAI-layout NVFP4 | Enhance LibertAI stock for the Spark vLLM recipe |
Abliteration does not change FLOPs. The intended speed win is shorter, more direct completions — not a different kernel path. Measured C1 on this dest: 22.3 tok/s, same band as stock LibertAI on this recipe.
Full credit to dealignai / @dealignai (compute @jordanschenck) for GLM-5.3-Flash-UNCENSORED-NVFP4.
Their NVFP4 packing is layout-identical to LibertAI (weight_map / tensor count / ModelOpt ignore). We byte-copied BF16 self_attn.o_proj for L15–45 onto LibertAI shards so the Spark vLLM kernels stay on the LibertAI expert packing. We did not ship their full checkpoint as a swap.
LibertAIDAI/GLM-5.3-Flash-NVFP4 is the parent: ModelOpt 0.45 weight-only NVFP4 on routed-expert FFNs, attention / vision / shared / MTP-as-quantized / embeddings BF16, ~181 GiB, 120 shards. Experts, vision, QKV, embeddings, and L0–14 o_proj remain theirs.
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. If you want their L2–44 / MTP-BF16 derisked NVFP4, use their repo.
Glm5NextForConditionalGeneration, hybrid KDA+DSA, mHC, native MTP).glm5_next + day-0 vllm/vllm-openai:glm53-flash-arm64-cu130.-1 init; barrydeen GMU 0.85 floor.nvfp4_ds_mla KV path used on this fleet.Byte-copy Dealign o_proj into LibertAI shards. Not a λ-projection of a rank-1 V.
| Tensor | model.language_model.layers.{L}.self_attn.o_proj.weight (BF16; in ModelOpt ignore) |
| Layers | 15–45 (31 tensors, includes MTP layers.45) |
| Mean Δrel vs LibertAI | 0.126 (L44 is Dealign’s 0.74; we did not invent that) |
| Safety | L0–14 byte-identical to LibertAI stock |
| Experts | NVFP4 passthrough (even in rewritten shards) |
| Gate | 32/32 bypass, 0 refuse, raw vLLM |
Variation table (why not L10–35 / Blackfrost V): see GitHub METHOD.md. Artifacts: ABLIT_META.json, VARIATIONS.json.
| Path | Purpose |
|---|---|
model-*-of-00120.safetensors + index | Full NVFP4 checkpoint (LibertAI layout, L15–45 o_proj from Dealign) |
ABLIT_META.json | Edit stats / recipe fingerprint |
VARIATIONS.json | Refusal32 + tok/s table |
tokenizer.json / chat_template.jinja / processor | Unchanged from LibertAI / Z.ai |
config.json is LibertAI stock (index_topk 2048). The Keys Spark recipe overlays index_topk=2044 at serve time — do not treat this card’s config as the overlay.
bash
# after you agree to the gate (automatic approval)hf download drowzeys/keys-GLM-5.3-Flash-NVFP4-ablit-l15-45-anchorstock \--local-dir ~/models/GLM-5.3-Flash-NVFP4-ablit-l15-45-anchorstock
Serve with the same vLLM GB10 recipe as LibertAI stock (marlin MoE, native MTP, NVFP4 KV as in the Keys Spark image). Do not pass --quantization modelopt_fp4 — honor config.json ignore. GPU memory utilization on this fleet is capped at 0.85 unless you have an explicit exception (this 1M+MTP stack uses 0.87).
MIT, inherited from Z.AI GLM-5.3-Flash (also the LibertAI NVFP4 card). You must still comply with the Responsible Use gate above. Blackfrost’s and Dealign’s own cards describe their artifacts; we did not ship Blackfrost’s weights, and we only copied Dealign o_proj L15–45 into a LibertAI body.