GLM-5.3-Flash-TQ3_4S

GGUF quantization of zai-org/GLM-5.3-Flash using TQ3_4S — a TurboQuant 4-bit format (~4.0 BPW nominal) with per-tensor precision overrides — 2-bit routed experts, 6-bit attention & shared experts, with the MTP (NextN) head retained for native speculative decoding.

Files

Download all 4 shards into one directory and point llama-server at shard 1.

File	Description
`tq3_4s/GLM-5.3-Flash-TQ3_4S.gguf-00001-of-00004.gguf`	Shard 1 (27.6 GiB)
`tq3_4s/GLM-5.3-Flash-TQ3_4S.gguf-00002-of-00004.gguf`	Shard 2 (27.8 GiB)
`tq3_4s/GLM-5.3-Flash-TQ3_4S.gguf-00003-of-00004.gguf`	Shard 3 (27.8 GiB)
`tq3_4s/GLM-5.3-Flash-TQ3_4S.gguf-00004-of-00004.gguf`	Shard 4 (18.5 GiB)
Total	101.6 GiB · ~4.0 BPW

Text-only artifact. The upstream base is image-text-to-text, but this quant ships no vision tower — there is no mmproj to load.

Quantization

MoE experts tolerate aggressive compression because only 8 of 288 are active per token. This quantization exploits that asymmetry:

Component	Quant	Rationale
Routed expert MLP	`iq2_s` / `q2_k`	2-bit, MoE-tolerant (8/288 active)
Shared experts + attention	`q6_K`	quality anchor
Norms / biases	`f32`	precision-sensitive
Embeddings + output	`q8_0`	quality anchor
`attn_output`	`q8_0`	keep the residual path clean
`blk.45` experts	`q2_k`	no imatrix coverage (documented workaround)
MTP / NextN head (`blk.45.nextn.*`)	`q6_K`	keeps the draft accurate for speculative decode

Requantized from unsloth/GLM-5.3-Flash-GGUF Q8_0 shards with the Unsloth imatrix (covers blk.3–44).

Runtime Requirement

This model requires the public TurboQuant runtime fork (mainline llama.cpp does not support the TQ3_4S tensor type):

https://github.com/turbo-tan/llama.cpp-tq3 (build ≥ 10403, CUDA sm_120/sm_121 or CPU)

Recommended Settings (single GB10 / DGX Spark, 121 GB unified)

LLAMA_MTP_PROCESS_ONLY=1 llama-server \
  -m tq3_4s/GLM-5.3-Flash-TQ3_4S.gguf-00001-of-00004.gguf \
  --spec-type draft-mtp --spec-draft-n-max 1 \
  -np 1 -c 40960 -ub 8 -b 8 -fa on -lm none -ngl 47 \
  --jinja

--jinja is required (custom chat template).
LLAMA_MTP_PROCESS_ONLY=1 keeps the draft KV correct under deep context (without it, long-context MTP can hang).
Context: native spec is 1M; on one GB10 with -np 1 we validated serving up to 512K. -c 40960 is the benchmarked shipping config.
The blk.45.nextn.* MTP head enables native speculative decoding — no separate draft model needed.

Performance (NVIDIA GB10 / DGX Spark, 121 GB, full clock)

Metric	Value
Decode (prose, temp 0.7, n=256, median of 3)	19.5 tok/s
MTP acceptance (dn=1)	100% deterministic · 87–90% prose
Size	101.6 GiB
BPW	~4.0 (nominal)
ngl	47 (unified-memory box)
Context (shipping)	40960 (validated to 512K)

Fits and serves on a single GB10 (121 GB unified) — no multi-GPU split needed.

Quality

Measured on the hardware above, greedy / temp 0.7 as noted.

Suite	Score
Hard86 (unit-test-pass, 86 coding tasks)	77.9% (67/86)
Coding (internal quality suite)	81.2 (9/12)
Toolcall	90.0 (13/15)
Data extraction	90.8 (12/15)
Instruction following	90.0 (13/15)
Reasoning / math	73.3 (11/15)
Quality composite	85.1

Base Model

zai-org/GLM-5.3-Flash (MIT)
Source: unsloth/GLM-5.3-Flash-GGUF Q8_0 (credit: Unsloth for the GGUF conversion and calibration/imatrix data)

License

Tool Call Validation

Tested with --jinja under both --reasoning off and --reasoning on --reasoning-budget 2048. Tool-call aggregate 13/15 (90.0): correct tool selection from multiple candidates, no spurious calls on plain questions, and tool-response → final-answer without runaway loops.

Verify against your own server:

# Start the server (see Recommended Settings), then:
curl -s http://127.0.0.1:8085/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model":"GLM-5.3-Flash-TQ3_4S","messages":[{"role":"user","content":"What is 27*43? Use the calculator tool."}],
       "tools":[{"type":"function","function":{"name":"calculator","description":"Eval math",
       "parameters":{"type":"object","properties":{"expression":{"type":"string"}},"required":["expression"]}}}]}'

YTan2000/GLM-5.3-Flash-TQ3_4S