Qwen3.8 Flash Next — TQ3_4S

TurboQuant GGUF build of Qwen3.8 Flash Next (qwen4exp architecture, 176.9B total parameters, MoE with shared experts, 48 layers, 256K context). Quantized to TQ3_4S (~4 bpw): the full model is 110 GiB across 5 shards and fits in the 128 GB unified memory of a single NVIDIA GB10 (DGX Spark class) machine.

Text-only conversion. The official base (and unsloth's F16 source) carry a vision tower and MTP draft layers; this GGUF quantization contains neither (verified at tensor level: 0 vision/mmproj tensors, 0 nextn/mtp tensors). No mmproj file is shipped and none is needed — vision inputs are not supported by this artifact, and no drafter is loadable with it.

Required Runtime

This release requires the TurboQuant runtime fork:

• https://github.com/turbo-tan/llama.cpp-tq3

Stock llama.cpp builds cannot load the custom TQ3_4S tensor type. The build must also include qwen4exp architecture support (merged August 2026 or later).

Files

Download all five shards into the same directory, then point llama-server at shard 1.

Recommended Runtime (validated)

Validated on one NVIDIA GB10 (128 GB unified):

llama-server \
  --model Qwen3.8-Flash-Next-TQ3_4S.gguf-00001-of-00005.gguf \
  -ngl 99 -c 32768 -fa on \
  --jinja \
  --reasoning-format deepseek --reasoning-budget 16384 \
  --port 8097 --host 127.0.0.1

• --jinja is mandatory (custom chat template).
• --reasoning-budget is mandatory for agentic loops — this is a reasoning model; an uncapped chain will run away.
• The full 262144-token context was verified to load and serve on a single GB10 (-c 262144 -fa on, ~74 GB resident) — the hybrid linear-attention architecture keeps the KV cache small. Use -c 32768 if you want headroom for concurrent slots.
• Measured decode: ~29.6 tok/s median (GB10, 32K ctx, no drafter).

Quick Smoke Test

curl -s http://127.0.0.1:8097/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "Qwen3.8-Flash-Next-TQ3_4S.gguf-00001-of-00005.gguf",
  "messages": [{"role": "user", "content": "Capital of France? One word."}],
  "max_tokens": 64, "temperature": 0
}'

Benchmark Summary

All numbers measured on the shipping artifact (GB10, TQ3_4S, 32K ctx, reasoning on with capped budget, official scorers):

Protocol notes: Hard86 and EvalPlus ran with the official scorers (reasoning on, budget 16384, 32K ctx). The quality-suite rows were re-run 2026-09-11 with thinking disabled, 24576-token budget, and a transport timeout sized for full-budget generations (an earlier run with a 600 s timeout and 8k budget truncated reasoning models mid-thought and is not comparable).

Parent Model & Provenance

Official base: Qwen/Qwen3.8-Flash-Next (qwen4exp architecture, 48 layers, 262144-token context).

Conversion lineage: this GGUF was quantized from unsloth's F16 GGUF (unsloth/Qwen3.8-Flash-Next-GGUF) using unsloth's imatrix calibration (imatrix_unsloth.gguf, 926 entries / 45 chunks), then re-quantized to the TurboQuant TQ3_4S type. The shipped GGUF metadata records this directly: general.quantized_by = Unsloth, general.repo_url = https://huggingface.co/unsloth, quantize.imatrix.file = Qwen3.8-Flash-Next-GGUF/imatrix_unsloth.gguf. unsloth's F16 GGUF is itself a conversion of the official Qwen/Qwen3.8-Flash-Next. Full recipe record: ai_workspace/recipes/flashnext-tq3_4s.yaml.

License

Qwen Community License 1.0 (inherits the official base model's license — the Qwen/Qwen3.8-Flash-Next LICENSE file — through unsloth's F16 GGUF, which we re-quantized).