scorpoon/Swift-Qwen3.8-Flash-Next-exl3-3.25bpw

scorpoon

~/quant-forge $ swift

Swift-Qwen3.8-Flash-Next · EXL3

independent exl3 build of ukisai/Swift1.5-Qwen3.8-Flash-Next · quantized from BF16 with the ExLlamaV3 converter

moe 262K context vision tower @ 6 bit MTP draft @ 4 bit n-gram table 5-bit shipped head 5 · mul1 · hq

downloads

The build lives on its own branch, following the layout of turboderp/Qwen3.8-Flash-Next-exl3. Pull it with --revision 3.25bpw_hq_ng5.

builds

branch	layer bpw	head	weights	requires
3.25bpw_hq_ng5	3.25	5.0	52.47 GiB · 7 shards	ExLlamaV3 ≥ v1.5.1

hq = high-quality head at 5 bits · metadata bpw 3.28 · MTP head 4-bit · vision tower 6-bit · the build ships its own 30.40 GiB 5-bit n-gram table (ng5), which lives in system RAM, not in the VRAM figure.

measurements · KLD

build	KLD mean	KLD p90	perplexity	Δ ppl
noise floor (BF16 vs BF16)	0.021	0.042	—	—
Swift BF16 (reference)	—	—	3.9453	—
3.25bpw hq ng5	0.107	0.214	3.9493	+0.10%

Evaluation trace disjoint from the calibration data · reference = Swift-Qwen3.8-Flash-Next in BF16, the direct base of this quant, not the upstream Qwen model · median KLD 0.023 · the noise floor is BF16 measured against itself and bounds what the method can resolve.

runtime · 2x RTX 5090 32 GB

build	MTP decode	draft acceptance
3.25bpw hq ng5	164–241 tok/s across eval categories, roughly 2x over MTP-off on coding tasks · peak ~48.3 GiB weights across both cards	1.86–3.49 draft tokens per step · at length 4: coding 74–92%, agentic/code 77–100%, creative 39–49%, translation 53–60%

The 52.47 GiB of weights exceed a single 32 GB card; the tested configuration autosplits across two GPUs with the n-gram table in system RAM. Single-card operation with CPU offload is possible in principle but was not tested and should cost decode speed. Stable config: chunk_size 3072, draft_num_tokens 4, dynamic_draft true.

how it was made

ExLlamaV3 v1.5.1 convert.py, calibration 250 x 2048 rows, codebook mul1, flags -b 3.25 -hb 5 -mb 4 -vb 6 -hq, separate 5-bit n-gram table (-ngf).

Only the weights were requantized and EXL3 metadata added; LICENSE, LICENSE-QWEN and NOTICE are the unmodified upstream files. Raw metrics (qb_results.json, draft_stats.json) live under eval/ on the build branch.

---

Original model card

Unmodified README of Swift1.5-Qwen3.8-Flash-Next, kept below for reference.

Website  •  Learn more  •  GGUF  •  GSQ-RCO GGUF  •  Evaluation  •  Enterprise licensing

Swift 1.5 Qwen3.8-Flash-Next

Swift 1.5 Qwen3.8-Flash-Next is UkisAI's reasoning-efficient derivative of Qwen3.8-Flash-Next. It uses 63.4% fewer thinking tokens, with a 1.8x speed up while keeping the accuracy loss <1% vs base on xhigh.

Demo

We gave base Qwen3.8-Flash-Next and Swift 1.5 Qwen3.8-Flash-Next the same prompt:

Create a 3D endless runner that has the fast, playful feel of Subway Surfers, but make the world and characters your own. I want to run through a lively place, dodge things, collect rewards, and feel the pace build the longer I survive. Make it fun to control and visually memorable. Use your judgment for the setting, mechanics, and little details that make it feel like a real game. Build it so I can launch and play it locally, then tell me how to run it.

Try the game yourself here: https://ukisai.com/swift-games/flash-next

Base Qwen3.8-Flash-Next took 8 minutes 52 seconds to build its game. Swift 1.5 took 4 minutes 56 seconds.

Training approach

We made Swift Flash Next efficient by figuring out which tokens were linked to pathological overthinking and penalizing them without "attacking" the reasoning length directly then regained the accuracy with RL and OPD, leading to "compressed" token usage while maintaining accuracy.

Swift 1.5 produces shorter reasoning traces. In our testing, we also observe fewer overthinking errors.

This release also features our previously mentioned post-training methods adapted specifically for coding and long-horizon agent work such as personal agents, terminal use and software engineering.

Our training data is viewable here: https://huggingface.co/datasets/ukisai/Qwen3.8-27B-multi-turn-agent-sft albeit it is not used out of the box, but rather re-sampled, turned into proper RL environments etc.

Evaluation

All scores compare the Qwen3.8-Flash-Next BF16 base with the Swift 1.5 BF16 checkpoint. Token columns report thinking tokens, except Terminal-Bench 2.1, which reports total generated output tokens.

Benchmark	Score		Mean tokens			Median tokens
	Base	Swift 1.5	Base	Swift 1.5	Reduction	Reduction
General reasoning
GPQA-Diamond	89.80%	89.60%	17,683	7,823	↓ 55.8%	↓ 63.4%
MMLU-Pro	87.75%	87.20%	3,528	1,519	↓ 57.0%	↓ 24.0%
C-Eval	93.27%	93.60%	1,048	586	↓ 44.1%	↓ 7.1%
IFBench	73.20%	70.13%	8,310	4,411	↓ 46.9%	↓ 55.6%
Mathematics
AIME 2026	98.67%	96.67%	23,015	15,806	↓ 31.3%	↓ 51.1%
HMMT (Nov 2025)	98.00%	97.33%	25,487	16,530	↓ 35.1%	↓ 54.7%
Multimodal
ERQA	70.80%	69.30%	4,036	1,788	↓ 55.7%	↓ 47.7%
Coding
LiveCodeBench v6	88.40%	90.39%	17,833	9,849	↓ 44.8%	↓ 52.0%
Agentic coding
Terminal-Bench 2.1*	67.64%	69.66%	40,591	45,428	↑ 11.9%	↓ 17.9%

* Note: Terminal-Bench 2.1 results for Flash-Next are not an apples-to-apples comparison with Swift 1.5 27B. Flash-Next used default task timeouts, leading to lower scores for both the base model and Swift Flash-Next.

How to reproduce

Serving: BF16 · Qwen3 reasoning parser · context 262,144 · thinking xhigh · MTP disabled.
Sampling: temperature 1.0 · top_p 0.95 · top_k 20 · min_p 0 · presence_penalty 0 · repetition_penalty 1.
Benchmarks: five seeds (0–4) for the seeded question benchmarks; Terminal-Bench 2.1 uses five attempts per task. LiveCodeBench is full release v6 mean pass@1 over seeds.
Terminal-Bench 2.1: Harbor 0.20.0 / Terminus-2 2.0.0, pinned 89-task dataset, JSON parser, interleaved thinking, temperature 1, top_p 1, 131,072-token server context, 3,600-second LLM call timeout and native per-task limits. Swift used concurrency 8; its context-recovery fix was applied during the run.

Benchmark	Output cap
GPQA-Diamond	100,000
MMLU-Pro	100,000
C-Eval	16,384
IFBench	81,920
AIME 2026	250,000
HMMT Nov 2025	250,000
ERQA	100,000
LiveCodeBench v6	100,000

Efficiency across reasoning efforts

GPQA-Diamond at each reasoning_effort setting, Swift 1.5 against the base at the same setting:

Reasoning effort	Score		Mean tokens			Median tokens
	Base	Swift 1.5	Base	Swift 1.5	Reduction	Reduction
Xhigh	89.80%	89.60%	17,683	7,823	↓ 55.8%	↓ 63.4%
Medium	86.36%	83.74%	4,157	2,483	↓ 40.3%	↓ 25.1%
Low	87.17%	84.75%	3,966	2,645	↓ 33.3%	↓ 19.7%

At xhigh, Swift 1.5 trails base by 0.20 percentage points while using 55.8% fewer mean and 63.4% fewer median thinking tokens. Medium and low save tokens but also lose 2.62 and 2.42 percentage points respectively.

Quantized models

Format	Repository	Runtime
AWQ INT4 (W4A16)	Swift-1.5-Qwen3.8-Flash-Next-W4A16-AWQ	vLLM (compressed-tensors)
AutoRound INT4 (W4A16)	Swift-1.5-Qwen3.8-Flash-Next-W4A16-AutoRound	vLLM (auto-round)
NVFP4	Swift-1.5-Qwen3.8-Flash-Next-NVFP4	NVIDIA Blackwell
GGUF	Swift-1.5-Qwen3.8-Flash-Next-GGUF	llama.cpp
GSQ-RCO GGUF (compact 2–3 bit)	Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF	llama.cpp

How to use

The BF16 checkpoint is about 360 GB, so it needs a multi-GPU node. We evaluated it with tensor parallelism 8.

GGUF download

The GGUF version is available for compatible llama.cpp-based runtimes. For the smallest files, use the GSQ-RCO GGUF version: 66.5–76 GB mixed-precision quants refined for Swift 1.5 Flash-Next.

UkisAI API

Swift 1.5 Flash-Next is served through an OpenAI-compatible API at https://ukisai.com/api/flash-next/v1. It is free for research purposes and needs no API key. The model id is flash-next.

from openai import OpenAI

client = OpenAI(base_url="https://ukisai.com/api/flash-next/v1", api_key="none")
response = client.chat.completions.create(
    model="flash-next",
    messages=[{"role": "user", "content": "Explain speculative decoding in two sentences."}],
)
print(response.choices[0].message.content)

curl https://ukisai.com/api/flash-next/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "flash-next", "messages": [{"role": "user", "content": "Hello, Swift."}]}'

vLLM

Use a vLLM build with Qwen3.8-Flash-Next support:

vllm serve ukisai/Swift-Qwen3.8-Flash-Next \
  --dtype bfloat16 \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_xml \
  --trust-remote-code \
  --port 8000

Reasoning effort is set per request through the chat template, for example "chat_template_kwargs": {"reasoning_effort": "low"}. xhigh is the default; medium and low are also supported.

SGLang

Alternatively, use an SGLang build with Qwen3.8-Flash-Next support, such as the lmsysorg/sglang:qwen38flashnext image:

sglang serve \
  --model-path ukisai/Swift-Qwen3.8-Flash-Next \
  --tp 8 \
  --context-length 262144 \
  --mem-fraction-static 0.85 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --port 8000

Adjust tensor parallelism and context length to your GPU memory. See the base model's vLLM recipe and SGLang cookbook for installation and hardware-specific settings.

Transformers

Use a Transformers release with Qwen3.8-Flash-Next support:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "ukisai/Swift-Qwen3.8-Flash-Next"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

To extend the context to 1,000,000 tokens, follow the YaRN settings on the Qwen3.8-Flash-Next model card.

Optional MTP decoding

The checkpoint includes the base model's one-layer MTP head. To enable self-speculative decoding, append the corresponding flags to the server command above:

# vLLM
--speculative-config '{"method":"mtp","num_speculative_tokens":1}'

# SGLang
--speculative-algorithm NEXTN --speculative-num-steps 3 \
  --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

The benchmark scores above were measured without MTP.

License and access

Swift 1.5 Qwen3.8-Flash-Next is a derivative of Qwen3.8-Flash-Next (Copyright (c) 2026 Qwen, Qwen Community License 1.0). UkisAI's contribution, including the adapted weights, is licensed under the Swift Open License v1.0. See NOTICE for the change notice and attribution details.

Personal, research, educational, evaluation, and commercial use of the Swift contribution are free for individuals and organizations with gross annual revenue, including affiliates, of up to US$1,000,000. Above that threshold, commercial use requires a separate Swift Enterprise License. Contact UkisAI for terms.

The base model's terms still apply to it. Under the Qwen Community License 1.0, organizations that run a Model-as-a-Service or AI Work Assistant business need a separate license from Qwen before any commercial use, and products above 100 million monthly active users or US$20 million monthly revenue must prominently display the model name. Nothing in the Swift Open License limits your rights in Qwen3.8-Flash-Next itself under the Qwen Community License.

Citation

@misc{swift-qwen3.8-flash-next,
  title  = {Swift 1.5 Qwen3.8-Flash-Next},
  author = {UkisAI},
  year   = {2026},
  url    = {https://huggingface.co/ukisai/Swift-Qwen3.8-Flash-Next}
}

Acknowledgements

We acknowledge the NVIDIA Innovation Lab, Amazon Web Services, and Google Cloud for providing compute credits and infrastructure support for Swift's development, training, and evaluation.