license: mit
base_model: orcarouter/GLM-5.3-Flash-Uncensored-FP8
base_model_relation: quantized
pipeline_tag: text-generation
library_name: transformers
inference: false
language:

• en
• zh
  tags:
• glm
• glm5
• glm-5.3-flash
• moe
• nvfp4
• compressed-tensors
• safetensors
• reasoning
• function-calling
• long-context
• uncensored
• abliterated
• not-for-all-audiences
• vllm

GLM-5.3-Flash OrcaRouter Uncensored NVFP4

ericlewis/GLM-5.3-Flash-OrcaRouter-Uncensored-NVFP4 is a mixed-precision,
weight-only NVFP4 conversion of
orcarouter/GLM-5.3-Flash-Uncensored-FP8,
pinned at source revision
3cec42d6ed14ec197e328c09650c17fd3660c26a.

The public release is published under the Hub tag v1.0.0. For a strictly
immutable reference, pin the commit SHA to which that tag resolves.

The official zai-org/GLM-5.3-Flash
BF16 checkpoint exists. This conversion deliberately starts from the pinned
OrcaRouter FP8 derivative so that its edited uncensored/abliterated lineage is
preserved. Re-quantizing the official BF16 checkpoint would produce a different
model.

What was converted

This is not an all-tensor NVFP4 checkpoint. The conversion uses a serving layout
that was validated with vLLM on NVIDIA DGX Spark:

The model-free converter streams one safetensors shard at a time. It does not
materialize the complete model or a complete BF16 intermediate. No calibration
dataset was used.

The source checkpoint contains 62 weight shards, 76,108 indexed tensor names,
and 37,338 block-FP8 modules. The converted checkpoint contains 62 weight
shards and 112,210 tensor names. Packed NVFP4 weights and their per-block and
global scales account for the larger output tensor count.

See CONVERSION.md for the exact partition, validation contract,
software versions, and checksums. The exact converter and validation programs
used for this artifact are included under conversion/.

Intended use

This release is intended for research, local inference, and evaluation of the
OrcaRouter-edited GLM-5.3-Flash lineage on NVFP4-capable NVIDIA hardware. The
qualified path exposes an OpenAI-compatible chat-completions API and supports
reasoning separation and structured tool calls.

This is an uncensored/abliterated derivative. It may follow requests that other
models refuse, and it can produce harmful, offensive, incorrect, or otherwise
unsafe content. Deployers are responsible for authentication, authorization,
rate limits, logging, output handling, and any application-specific safeguards.
Do not expose an unauthenticated inference port to an untrusted network.

The source model card reports safety evaluations for the FP8 derivative. Those
evaluations were not rerun for this NVFP4 conversion, and quantization can alter
model behavior.

Tested runtime

The checkpoint passed qualification with the following configuration:

The exact qualification runtime was rebuilt from the public base with an SM121
sparse-indexer replacement and a vLLM prefix-cache correction. That derived
runtime image is not part of this publication. The included public SparkRun
recipe is therefore a topology-neutral reference and must be requalified on the
operator's runtime; it is not claimed to reproduce the private qualification
image byte-for-byte.

Qualification results

Functional and long-context checks

• Served-model identity: passed.
• Basic generation: passed.
• Reasoning separation: passed; the receipt recorded 196 reasoning characters.
• Strict final-answer formatting: passed.
• One structured tool call: passed.
• Multi-turn tool-result replay: passed.
• Parallel structured tool calls: passed.
• Correct no-tool decision: passed.
• Streaming structured-tool parsing: passed.
• UTF-8 stress probe: 30/30 responses passed with no replacement characters,
  malformed payloads, missing IDs, or incomplete tails.
• Long-context retrieval: passed at 122,578 prompt tokens and 111 completion
  tokens in 73.775 seconds, using an exact hidden-marker match.

The checkpoint config advertises max_position_embeddings: 1048576. Only the
131,072-token server configuration and the 122,578-token retrieval prompt were
qualified here. This publication does not claim a verified one-million-token
deployment.

Prefix-cache probe

The probe used an identical 4,045-token prompt:

Matched C1 performance screen

Each result below is the median of three measured 256-token rounds after one
warmup, with one running request. The FP8 control used the pinned OrcaRouter
source checkpoint and otherwise matched the serving screen.

The two checkpoints ran on different four-node subsets and the rounds were not
interleaved. These measurements describe one local serving stack, not a general
performance guarantee. This was a transport/performance screen: all measured
samples reached the 256-token limit and failed the benchmark harness's embedded
answer-quality checklist. The table is therefore speed evidence only. The
separate functional and long-context checks above provide the correctness
evidence.

Machine-readable details and receipt hashes are in
provenance/qualification-summary.json.

Independent replica and routed Hermes acceptance

The same checkpoint and runtime were then deployed as a second independent TP4
replica on four additional DGX Spark systems. The second replica repeated the
functional, reasoning, tool-call, 122,578-token retrieval, UTF-8, and prefix
cache gates successfully.

Its matched C1 screen measured 47.7020 output tok/s median, 262.596 ms median
TTFT, 316.283 ms p95 TTFT, and 0.70175 DFlash2 acceptance. The exact active
Cosmos prompt measured 33,753 tokens; its two warm requests each reused 32,256
cached prompt tokens.

Both replicas were then placed behind one cache-aware router with pool
concurrency two. Concurrent probes reached both workers. A real Hermes/Cosmos
request through the routed endpoint completed two model calls, kept 314 and 166
characters of private reasoning separate across the two assistant turns,
executed exactly one structured terminal tool call, returned the exact requested
marker, and reported 18,432 cached prompt tokens. The Hermes gateway, its
private SSH tunnel, and all eight model containers remained running without a
restart or OOM kill during acceptance.

SparkRun usage

The public reference recipe is
recipes/sparkrun-tp4-dgx-spark.yaml.
It pins this checkpoint to v1.0.0, pins the DFlash2 revision, and retains the
qualified TP4, KV-cache, parser, and scheduler settings without embedding any
site-specific host, interface, cache-snapshot, or registry coordinates. The API
defaults to loopback; override the host only with a trusted fabric address and
an explicit firewall policy.

Run it only on a trusted cluster network, and verify that the selected runtime
contains working SM121 kernels for this checkpoint before promoting it.

For direct vLLM use, the equivalent core arguments are:

vllm serve ericlewis/GLM-5.3-Flash-OrcaRouter-Uncensored-NVFP4 \
  --revision v1.0.0 \
  --tokenizer-revision v1.0.0 \
  --served-model-name glm-5.3-flash \
  --host 127.0.0.1 \
  --tensor-parallel-size 4 \
  --load-format safetensors \
  --quantization compressed-tensors \
  --moe-backend marlin \
  --dtype bfloat16 \
  --kv-cache-dtype fp8_e4m3 \
  --max-model-len 131072 \
  --max-num-seqs 1 \
  --max-num-batched-tokens 4096 \
  --kv-cache-memory-bytes 4294967296 \
  --block-size 2304 \
  --speculative-config '{"method":"dflash","model":"incoai/GLM-5.3-Flash-DFlash2","revision":"dc77ff1c99eeb2df044ee3d4f0094eb033fee410","num_speculative_tokens":3}' \
  --enforce-eager \
  --disable-custom-all-reduce \
  --no-enable-flashinfer-autotune \
  --enable-prefix-caching \
  --enable-prompt-tokens-details \
  --language-model-only \
  --trust-remote-code \
  --enable-auto-tool-choice \
  --tool-call-parser glm47 \
  --reasoning-parser glm45

The command requires a compatible vLLM build and enough aggregate memory. It
does not configure cluster networking or access control.

Artifact identity and integrity

The 64-character local conversion artifact ID is a deterministic local artifact
identifier, not a Hugging Face commit hash. Consumers should pin the public
v1.0.0 tag or the Hub commit to which that tag resolves.

The release checkpoint consists of config.json, tokenizer/configuration
assets, model.safetensors.index.json, and 62 files matching
model-*-of-00062.safetensors. Publish homebot-conversion-manifest.json and
this qualification summary alongside the checkpoint so downstream users can
audit the conversion contract. SHA256SUMS covers every release
payload file except itself.

Limitations and unverified configurations

• The output is mixed precision. Calling every tensor NVFP4 would be inaccurate.
• Official BF16 weights exist, but were intentionally not used because they do
  not contain the OrcaRouter edits.
• The MTP expert layer remains block FP8, and 186 other source FP8 matrices are
  BF16 in the output.
• The source model's safety-evaluation claims were not rerun after conversion.
• The retained vision configuration was not exercised; qualification used
  --language-model-only.
• Tensor-parallel sizes other than TP4, pipeline parallelism, data parallelism,
  per-replica concurrency above one, CUDA graphs, training, and fine-tuning
  were not qualified for this artifact. Pool-level concurrency two was tested
  with two independent C1 TP4 replicas.
• The one-million-token configured limit was not tested.
• The public base image and public recipe were not rerun as a byte-identical
  replacement for the derived qualification runtime.
• Results may change with vLLM, kernel, driver, firmware, interconnect, or
  drafter revisions.

Licenses

The model checkpoint inherits the MIT license distributed with the OrcaRouter
source and the Z.AI base model. The optional
incoai/GLM-5.3-Flash-DFlash2
drafter is separately licensed under CC BY-NC-ND 4.0. Using that drafter makes
the combined serving configuration subject to its non-commercial restrictions;
the checkpoint can be served without DFlash2.

Acknowledgements

This conversion builds on GLM-5.3-Flash from Z.AI, the OrcaRouter edited FP8
derivative, compressed-tensors/llmcompressor, vLLM, and the IncoAI DFlash2
drafter.