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.
115 engagement1 sourceWed, Sep 16, 2026, 11:09 AM
AI Overview
- ISTA-DASLab's GSQ-RCO quants shrink Qwen3.8-Flash-Next from ~80-95GB to 68-76GB while preserving near baseline quality
- Q2_0 variant delivers 3.4x prompt throughput and 1.9x lower latency than IQ2_XS, scoring 89.07 vs IQ2_XS's 89.16 task average
- IQ3_XXS matches the base model exactly on AIME25 (100.00), within 0.51 on GPQA-Diamond and 1.14 on LiveCodeBench v6 at ~one fifth BF16 size
Related
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.
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 sourceredditTue, Sep 29, 2026, 08:40 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
Qwen 3.8 flash for a single spark
A community member released a low-bit quantization of the Qwen 3.8 Flash base model on Hugging Face. The quant reportedly retains 95% of bfloat16 accuracy without degradation at long contexts and reaches roughly 40 tokens/s, leaving headroom for 256k context and RoPE. Inference improvements are still in progress.
A community member released a low-bit quantization of the Qwen 3.8 Flash base model on Hugging Face. The quant reportedly retains 95% of bfloat16 accuracy without degradation at long contexts and reaches roughly 40 tokens/s, leaving headroom for 256k context and RoPE. Inference improvements are still in progress.
3 engagement1 sourceredditFri, Oct 2, 2026, 11:27 AM
Social·Community
Qwen3.8-Flash-Next IQ3_S on Strata via pi scores 49/49 on airbench in 11 min on an RTX 5090, beating Claude Code's 14 min
A Reddit user reports running the ISTA-DASLab IQ3_S quant of Qwen3.8-Flash-Next via Strata and the pi coding agent on a single RTX 5090, scoring 100% on the airbench everyday-tasks suite in 11 minutes versus 14 minutes for Claude Code Opus 5.5. The post includes full reproduction instructions and comparison runs from other harnesses on the same setup. The author notes the original title was misleading since airbench targets everyday tasks, not frontier-model discrimination.
A Reddit user reports running the ISTA-DASLab IQ3_S quant of Qwen3.8-Flash-Next via Strata and the pi coding agent on a single RTX 5090, scoring 100% on the airbench everyday-tasks suite in 11 minutes versus 14 minutes for Claude Code Opus 5.5. The post includes full reproduction instructions and comparison runs from other harnesses on the same setup. The author notes the original title was misleading since airbench targets everyday tasks, not frontier-model discrimination.
44 engagement1 sourceredditSat, Oct 10, 2026, 09:54 PM
Releases·Tool Release
Community member releases imatrix GGUF quants for Qwen3.8-Flash-Next, 20-30GB smaller than Unsloth builds
A community quantizer released mainline-compatible imatrix GGUF quants for Qwen3.8-Flash-Next on Hugging Face after several days of testing. The builds target the same quality band as popular Unsloth and AesSedai Q4 quants while saving roughly 20-30GB of disk and RAM, with perplexity competitive with both. This gives local inference users smaller files without a quality tradeoff.
A community quantizer released mainline-compatible imatrix GGUF quants for Qwen3.8-Flash-Next on Hugging Face after several days of testing. The builds target the same quality band as popular Unsloth and AesSedai Q4 quants while saving roughly 20-30GB of disk and RAM, with perplexity competitive with both. This gives local inference users smaller files without a quality tradeoff.
39 engagement1 sourceredditSat, Aug 29, 2026, 03:02 PM