Releases·Model Release
ISTA-DASLab releases GSQ-RCO GGUFs for Qwen3.8-Flash-Next plus 50% expert-pruned Coder build at ~1.89 bpw
ISTA Deep Algorithms and Systems Lab released GGUF quantizations of the 176.9B-parameter sparse MoE model Qwen3.8-Flash-Next using two new methods, GSQ and RCO. Four quantized builds run from 2.40 to 3.50 bpw, and a Coder build removes half the routed experts, yielding an effective ~1.89 bits per original parameter and a 29.6 GB resident footprint that fits a single 32 GB accelerator. Benchmarks show the 3.50 bpw build matching BF16 on every evaluated benchmark.
164 engagement1 sourceTue, Sep 29, 2026, 08:40 AM
AI Overview
- ISTA-DASLab releases Qwen3.8-Flash-Next GGUFs at 2.40-3.50 bpw (66.4-83.6 GB); IQ3_S matches BF16, task average 93.26 vs 93.12
- RCO selects 256 of 512 experts per layer for the Coder build; retained weights stay at 3.5 bpw
- Pruned Coder build averages 1.89 bpw over the original 176.9B parameters, fitting 29.6 GB resident on a single 32 GB accelerator
- GSQ learns grid assignments and group scales post-training, closing the scalar-vector quantization gap in standard GGUF types
- Pruned Coder scores SWE-bench Verified 75.60 (91.3% of BF16) and LiveCodeBench v6 86.28 (98.7%) at xhigh reasoning effort
- Q2_0 at 2.40 bpw (66.4 GB) scores zero-shot average 78.00 vs BF16's 76.94 at roughly one fifth the size
Related
Releases·Tool Release
ISTA-DASLab releases GSQ-RCO GGUFs for Qwen3.8-Flash-Next: near-baseline quality at 68-76GB, up to 3.4x prompt throughput
ISTA-DASLab published GSQ-RCO quantized GGUFs of Qwen3.8-Flash-Next on Hugging Face, shrinking the model from ~80-95GB to 68-76GB while preserving near baseline quality. The release targets practitioners running the model locally at reduced memory cost.
ISTA-DASLab published GSQ-RCO quantized GGUFs of Qwen3.8-Flash-Next on Hugging Face, shrinking the model from ~80-95GB to 68-76GB while preserving near baseline quality. The release targets practitioners running the model locally at reduced memory cost.
115 engagement1 sourceredditWed, Sep 16, 2026, 11:09 AM
Releases·Tool Release
Two new quantization releases target Qwen3.8-27B: GSQ-RCO GGUFs and QUASAR NVFP4
Two independent teams released aggressively quantized versions of Qwen3.8-27B within two days. One ships GGUF checkpoints at 2.5-3.0 bits per weight using two new methods, GSQ and RCO, with both the search and quantizer learned. The other releases an NVFP4 W4A4 checkpoint trained via quantization-aware distillation with the QUASAR QAT algorithm, deployable on vLLM with NVIDIA Blackwell GPUs. Both matter to practitioners running large models at low precision locally or on modern accelerators.
Two independent teams released aggressively quantized versions of Qwen3.8-27B within two days. One ships GGUF checkpoints at 2.5-3.0 bits per weight using two new methods, GSQ and RCO, with both the search and quantizer learned. The other releases an NVFP4 W4A4 checkpoint trained via quantization-aware distillation with the QUASAR QAT algorithm, deployable on vLLM with NVIDIA Blackwell GPUs. Both matter to practitioners running large models at low precision locally or on modern accelerators.
Updated Fri, Aug 28, 2026, 09:46 PM — GSQ-RCO GGUF release at 2.5-3.0 bpw followed the earlier QUASAR NVFP4 checkpoint on 2026-08-28.
141 engagement2 sourcesredditWed, Aug 26, 2026, 01:04 AM
Social·Community
Running Qwen3.8-Flash-Next, a 125B MoE model, on one RTX 3090 Ti
Developer runs Qwen3.8-Flash-Next 125B MoE on a single RTX 3090 Ti
Developer runs Qwen3.8-Flash-Next 125B MoE on a single RTX 3090 Ti
0 engagement1 sourcerssSun, Oct 4, 2026, 01:25 AM
Releases·Tool Release
Strata v0.1.40 adds experimental Qwen3.8-Flash-Next support on Strix Halo, hitting 1M context
Developer Niko1221 released Strata v0.1.40 with official support for running Qwen3.8-Flash-Next on Strix Halo machines, claiming best-in-class long-context decode and prompt throughput using Unsloth Q4 and GSQ-RCO weights. It supports up to 1M context with little speed loss, but support is experimental and Linux-only.
Developer Niko1221 released Strata v0.1.40 with official support for running Qwen3.8-Flash-Next on Strix Halo machines, claiming best-in-class long-context decode and prompt throughput using Unsloth Q4 and GSQ-RCO weights. It supports up to 1M context with little speed loss, but support is experimental and Linux-only.
225 engagement1 sourceredditTue, Oct 6, 2026, 02:58 PM
Social·Community
Qwen 3.8 Flash-Next (176B MoE) runs at 500k context on a single AMD 7900 XTX — and a context-echo bug traced to preserve_thinking
A user's detailed writeup reports running Qwen3.8-Flash-Next on a single AMD 7900 XTX with 500k context, after diagnosing a 'reasoning echo' bug tied to preserve_thinking: true across nearly 13,000 turns. Measured results — 105–160 tok/s decode, quality holding 81–94% — position Flash-Next as a viable local daily driver on consumer AMD hardware.
A user's detailed writeup reports running Qwen3.8-Flash-Next on a single AMD 7900 XTX with 500k context, after diagnosing a 'reasoning echo' bug tied to preserve_thinking: true across nearly 13,000 turns. Measured results — 105–160 tok/s decode, quality holding 81–94% — position Flash-Next as a viable local daily driver on consumer AMD hardware.
118 engagement1 sourceredditSun, Oct 11, 2026, 09:44 PM