peasantsmith/Qwen3.8-Flash-Next-PS-IQ2_XXS-GGUF

Qwen3.8-Flash-Next PS-IQ2_XXS (Peasant Smith)

176.9B parameters in a single 75.2 GB file — 3.40 bits/param. An IQ2_XXS-class GGUF of Qwen3.8-Flash-Next, built from the original Qwen/Qwen3.8-Flash-Next BF16 weights, with every component quantized to the precision it needs instead of crushed flat.

Please read: After running it for days non-stop I have concluded that this quantization is pretty good, but it does require certain care. I recommend running it with Thinking: Disabled or Low. You can definitely use xHigh or Medium, but I won't recommend them on hours of uninterrupted agentic work since you may risk loops or simply wasting time on overthinking. For short/medium tasks using thinking as xHigh is completely fine.

How the size was won

The model's single largest component is a per-layer-embedding (PLE) n-gram table: 51.2B params (~102 GB in BF16) — roughly 28% of the model. A naive 2-bit quant crushes it along with everything else, making the model's biggest component its weakest. This build takes the 2-bit win where it's safe and holds the PLE table at the quality floor the ecosystem agrees on:

Model component	Params	Quant	Size
PLE n-gram table	51.2B	IQ4_NL (4.5 bpw)	26.8 GiB
Routed expert gate/up	—	IQ2_XXS (2.06 bpw)	19.3 GiB
Routed expert down	—	Q4_0 (4.5 bpw)	21.1 GiB
Attention (QKV)	—	Q4_K	0.5 GiB
Token embeddings	—	Q6_K	0.5 GiB
Output head	—	Q8_0	0.8 GiB
Routers / norms / biases / hash constants	—	F32 (kept)	—
Vision encoder (mmproj, 334 tensors)	—	F16	0.9 GiB

• Expert down-projections sit a full step above gate/up (Q4_0 vs IQ2_XXS): their errors land directly and additively on the residual stream, while gate/up errors pass through activations first.
• Every router, norm, bias, and hash-constant tensor stays F32 — the parts that must stay high precision aren't compressed at all.
• Imatrix: unsloth's published importance matrix for this model, audited to cover every quantized tensor (hyper-connection and PLE tensors pinned out of IQ quantization).

Testing

Test	Result
Full-context generation (64k context window)	✅ coherent at depth, no degradation
Long-generation stability	✅ ≥3000 tokens continuous at steady t/s, no degradation
Tool calling (get_weather JSON function call)	✅ well-formed {"city":"Paris"} tool call returned
Long-context recall (~6k token prompt, needle)	✅ number recalled correctly
LRU cache code task (7.5k token prompt)	✅ complete, correct implementation
Greedy smoke ("The capital of France is")	✅ "Paris" (finish=stop)
Vision via mmproj (image color ID)	✅ correct
Perplexity (wikitext-2)	not measured — honesty note below

Honesty note: PPL was not run — at this weight class the BF16 reference does not fit the same GPU stack, so no comparable baseline exists on this hardware. The task battery above is the quality evidence; 2-bit experts do carry a real quality delta vs BF16 by construction.

Files

File	Size	Notes
Qwen3.8-Flash-Next-IQ2_XXS.gguf	75.2 GB	text model (176.9B params)
mmproj-Qwen3.8-Flash-Next-F16.gguf	0.90 GB	vision encoder, F16, all 334 tensors — nothing stripped

Usage

Generic llama.cpp invocation — tune the two offload values to your own hardware:

llama-server \
  -m Qwen3.8-Flash-Next-IQ2_XXS.gguf \
  --mmproj mmproj-Qwen3.8-Flash-Next-F16.gguf \
  -ngl <layers that fit your VRAM> \
  --n-cpu-moe <experts to spill to CPU RAM if VRAM is tight> \
  --jinja

The model ships its chat template (--jinja). Requires a recent llama.cpp with qwen4exp architecture support.

Provenance & credits

1. Qwen/Qwen3.8-Flash-Next — official BF16 safetensors, original weights (no third-party quant or modified repo as source). Model credit: Qwen.
2. → GGUF master via llama.cpp convert_hf_to_gguf.py
3. → llama-quantize importance-matrix quant with the per-component split above. Imatrix credit: unsloth.

Verify

SHA-256 (text): 2e0f14e7eeddce8f80fc88cf96a9cc641b4f60549318fb53e30af85649883586 SHA-256 (vision): f5f125bdfa3997011693e7fe59fae1a5c068c40a627a0576d72734a9cdab9691

Limitations

• 2-bit experts: a real quality delta vs BF16 exists by construction — the task battery passed, but long-tail knowledge degradation is expected at this footprint
• Text model only: the vision tower ships as the separate F16 mmproj file
• Requires a recent llama.cpp (qwen4exp architecture support)
• MTP head not embedded — left as shipped by Qwen