Sakura Qwen3.8 Flash-Next Compact · ISTA IQ3_S · 512E · CP2

Experimental community repack of ISTA-DASLab GSQ-RCO IQ3_S, retaining all 512 routed experts. The original first GGUF shard is copied unchanged. Only the separate N-Gram/PLE lookup-table shard is replaced with the existing Sakura CP2 table.

This exact IQ3_S + CP2 package has not been loaded for inference or tested for generation, perplexity, quality or speed. Its pinned file hashes, GGUF split layout, lookup addressing and tokenizer metadata have been checked. Publication here is an experimental structural release, not a model-quality or hardware qualification.

Files and size

The model files total 68.258 GB in decimal download units. The source two-shard set totals 83.618 GB; replacing its IQ4_NL table saves about 15.360 GB. The small runtime and documentation are additional. File size is not GPU memory usage or proof that this configuration fits a particular machine.

The backbone remains ISTA's original IQ3_S mixed quantization. CP2 describes the separate experimental table representation; this is not a stock whole-model quantization preset. Internal main-shard metadata, including the source model name and original tokenizer/template, remain unchanged.

Download and start on Windows

Download the complete repository into one folder:

hf download webmp3/Sakura-Qwen3.8-Flash-Next-Compact-ISTA-IQ3_S-512E-CP2-GGUF --local-dir Sakura-IQ3_S-CP2

Keep both numbered GGUF files, START.cmd and the complete runtime directory together. Start START.cmd, wait for loading, then open http://127.0.0.1:18098. The OpenAI-compatible endpoint is /v1/chat/completions, with alias sakura-flash-next-iq3-s-cp2. Stop the server with Ctrl+C.

The proposed starter uses the included Windows x64 Vulkan CP2 runtime, one slot, context 4096, Q8 KV cache and lazy table reads from local NVMe. It preserves the source chat template and its thinking default. MTP, vision and the optional CP2 GPU row cache are disabled. This starter has been checked as a release file but has not been executed against this exact package.

Use the included runtime: stock llama.cpp, LM Studio, Ollama and unrelated forks are not qualified for this custom CP2 table. No AMD/64-GB-UMA fit, peak-memory, long-context or performance result is claimed for this variant. Settings may need adjustment for the user's actual available memory.

What was verified

• The original first shard is identified by its full HF LFS SHA256 and is copied without payload or metadata changes.
• The original second shard has the same full-file SHA256 as the common IQ4_NL input used by the existing Sakura CP2 repack, not merely the same tensor shape.
• The original IQ4_NL table file and the existing CP2 table file were also fully rehashed locally on 2026-10-21 and matched their recorded hashes. The CP2 table is a lossy derivative, not byte-identical to IQ4_NL.
• The first shard has 1,223 tensors; the replacement has one table tensor. Both use the two-shard contract for 1,224 tensors in total.
• The lookup-addressing metadata and complete raw tokenizer metadata match the checked ISTA Q2 reference.
• The existing CP2 runtime binaries and notices are reused unchanged from the pinned public Sakura K512 repository. No new encoder, training, expert pruning or runtime port was performed.

See results.json for scope and provenance, PROVENANCE.json for pinned sources and SHA256SUMS for all shipped files. Previously reported Sakura Q2 or Gyro-M scores are not results for this package and are not imported here. Model-file integrity after the server-side copy is checked against HF's LFS hash/size metadata; no independent full download of the new repository is performed.

Provenance and license

Base: Qwen/Qwen3.8-Flash-Next. Direct backbone: ISTA-DASLab GSQ-RCO, revision ed59f92082b1e93c0e96d60a8b11aab089b52f09. Existing CP2 shard and matching runtime: Sakura K512 CP2, revision 9b8fef370b812b3fd33787869d553793517245cb.

Credit for the GSQ-RCO quantized backbone belongs to ISTA-DASLab and its contributors; credit for Qwen belongs to Qwen. Sakura supplies the experimental lookup-table representation and this repack. This release claims no authorship of the backbone, no retraining and no endorsement by those projects.

Model weights retain Qwen Community License 1.0, included in LICENSE and LICENSE-QWEN. Read its complete conditions before use. The upstream HF card's Apache tag does not replace the inherited model-weight license here. Runtime software has its separate MIT and third-party license notices under runtime/; software licensing does not relicense model weights.