RESMP-DEV/Qwen3.8-Flash-Next-IQ2_XS-Q2_0-iMatrix-Calibrated-256x2048-GGUF

Qwen3.8 Flash-Next · IQ2_XS / Q2_0 iMatrix-Calibrated 256×2,048 GGUF

This repository contains a research static requantization of Qwen/Qwen3.8-Flash-Next. It is not a runtime-dynamic quantization and it is not a general-purpose release recommendation. The routed expert weights are fixed at IQ2_XS or Q2_0 when written to the GGUF; routing at inference remains dynamic as in the original mixture-of-experts model.

Status: quantized, integrity-checked, packed, and loaded by Strata on a real RTX 5070 Laptop host. A one-token greedy gate and a same-day public-iMatrix control both produced token 157. No perplexity, task-quality, long-generation, or broad throughput evaluation has been run.

Important scope and disclaimer

• The source was an already quantized IQ4_XS GGUF, not the original BF16 checkpoint. This artifact is therefore a compounded research requantization; it must not be interpreted as a quality-equivalent low-bit conversion from BF16.
• 96 routed gate/up tensors use IQ2_XS and 48 routed down tensors use Q2_0. The other 1,080 tensors retain the intermediate GGUF's existing types and shapes.
• Calibration used a private RESMP-DEV exact-token iMatrix view containing 256 windows of 2,048 tokens. The calibration corpus and importance matrix are not published with this model.
• The GGUF retains F32 router tensors. Strata's native pack path converts those routers exactly to BF16 while leaving expert and PLE bytes unchanged.
• This is not a default Strata model choice and has not been promoted beyond research.

At a glance

Property	Value
Architecture	Qwen3.8 Flash Next / qwen4exp mixture-of-experts
Runtime tested	Strata engine 0.1.41
Tensor count	1,224 across two llama.cpp GGUF shards
Routed gate/up experts	48 IQ2_XS tensors
Routed down experts	48 Q2_0 tensors
Non-target tensors	1,080 preserved source types and shapes
Unsharded artifact size	66,649,935,712 bytes
Reported quantized size	63,551.82 MiB / 3.01 bits per weight
Calibration view	256 disjoint windows of 2,048 token IDs
Deterministic IQ2_XS path	Enabled for the described research build

Files

These are a matching llama.cpp GGUF split set; neither shard is an independent model. Sizes below are decimal GB and describe files on disk, not runtime memory.

File	Size	SHA-256
Qwen3.8-Flash-Next-IQ2_XS-Q2_0-iMatrix-Calibrated-256x2048-00001-of-00002.gguf	39.969 GB	3c69f42a305c15209893f0e1c8b70ffcbc9dd8346cdd7c781e7edefbe42b2915
Qwen3.8-Flash-Next-IQ2_XS-Q2_0-iMatrix-Calibrated-256x2048-00002-of-00002.gguf	26.681 GB	adf8e9e382276a7ff20150c0f3701adb847c4a140895a693367c4f7400123036

The unsharded source artifact has SHA-256 f88ddbabfe3a7c307c4685ae524efddcceada6e5b427d0d5aa574a53a767cc08. Splitting and merging reconstruct GGUF split metadata and alignment, so the merged file need not have the same whole-file hash. Tensor-level verification is authoritative: all 1,224 names, shapes, types, byte counts, and payload hashes matched before upload.

Usage

Download both shards to the same directory. Keep the complete filenames together:

hf download RESMP-DEV/Qwen3.8-Flash-Next-IQ2_XS-Q2_0-iMatrix-Calibrated-256x2048-GGUF \
  --local-dir ./Qwen3.8-Flash-Next-IQ2_XS-Q2_0-iMatrix-Calibrated-256x2048-GGUF

A current llama.cpp llama-gguf-split --merge can reconstruct a single GGUF from shard 1. Use the exact binary and commands for your llama.cpp version; merge behavior was verified, but llama.cpp text generation was not part of this artifact's validation. Strata validation used the unsharded GGUF and Strata's native pack path.

Calibration and conversion

The input was mradermacher/Qwen3.8-Flash-Next-GGUF, file Qwen3.8-Flash-Next.IQ4_XS.gguf, pinned at revision 68d514cd9f411c9d6b62f8a422102a202d57c18c and SHA-256 ea9b7537389918d69bcf3ea1fb5693d427ae82278dd67f415fa9383858b51efb.

The RESMP-DEV calibration view contains 524,288 token IDs in 256 disjoint 2,048-token windows. Its resulting importance matrix has SHA-256 92c6a742027b569907a2ce3cd2bf12b3ae5ea037c86105ef7ba4ae244a43fc89 and 902 entries. The matrix itself is retained locally and is not required for inference.

Quantization used llama.cpp revision 3cf03257f219afbe7334045ff7c6a06ac68c627d with two opt-in research paths: double-precision deterministic IQ2_XS derived-weight arithmetic and a complete anchored per-tensor COPY recipe. The full run used 24 threads, took 27:48.85, and peaked at 29.33 GiB RSS.

Validation

The artifact's 1,224 tensor names and shapes match the source. Exactly 144 intended routed-expert conversions occurred: 96 gate/up tensors became IQ2_XS and 48 down tensors became Q2_0. No non-target tensor type changed.

Strata loaded 32.23 GiB of native experts at 6.93 GiB/s, selected 2,233 expert-cache slots on an RTX 5070 Laptop GPU, and generated greedy token 157 for prompt tokens [100,101,102]. Measured rates for that one-token gate were 26.33 tokens/s decode and 6.83 tokens/s prefill. A same-boot control using the prior public-matrix artifact, the same engine, and the same prompt also produced token 157 at 24.96 tokens/s decode and 5.45 tokens/s prefill.

These checks prove artifact integrity and execution, not quality. The speed comparison is a same-day observation, not an interleaved confidence interval. No perplexity, coding, reasoning, agentic, long-context, vision, or general throughput benchmark has been run.

Reproducibility and integrity

• Source GGUF: 98,421,788,160 bytes; SHA-256 ea9b7537…b51efb.
• Importance matrix: 579,789,056 bytes; SHA-256 92c6a742…a43fc89.
• Unsharded artifact: 66,649,935,712 bytes; SHA-256 f88ddbab…67cc08.
• Split/merge tensor verification: 1,224 matched payloads, 0 mismatches.
• Strata source revision: e19b1d73a08ff4205925b0bc5697c70ec1e6bded.

See SHA256SUMS and provenance.json for structured integrity data.

Credits

Original model: Qwen. Intermediate GGUF: mradermacher. Quantization runtime: ggml-org/llama.cpp. Native engine validation: Strata. Standard exact-token iMatrix view, deterministic research quantization, split verification, and publication: RESMP-DEV.