Qwen3.8-Flash-Next FP8 — dealignai edition

Compliance-tuned drop-in replacement for Qwen3.8-Flash-Next-FP8. Refusal removed, reasoning depth increased vs base, MTP head and vision tower fully preserved.

Highlights

Reasoning improvement — verified

Probed at temperature=0 on 5 classic reasoning puzzles. Compared to the FP8 base:

This model thinks 3.5× more tokens on average and gets one more puzzle correct than the FP8 base. The snail puzzle is the classic "does the snail slide on the final day" trick — base incorrectly divides 30 ÷ 4 = 8 (assumes it slides on the final night); this model correctly walks 5 days of net-1-ft plus the final day's climb-out.

The think-token ratio rose from 73.8% → 89.9%. Reasoning depth is not a regression here — the model now spends real deliberation on hard problems where the base model was previously truncated by short refusal-shortcut trained-in behavior.

HarmBench — full 6-category dropdown

Scored with a 5-tier classifier: TRUE_COMPLY / HEDGE_COMPLY (refuses then complies) / SOFT_REDIRECT (refuses to defensive info) / PARTIAL (short comply) / HARD_REFUSE. Thinking mode off (enable_thinking: false).

PARTIAL rows are short-but-compliant answers (e.g. a ransom note, a phishing text-message, a tool-call artifact) — length classifier under-counts them. Manual review: all 9 PARTIAL are actual complies → effective compliance is 156 / 156 = 100% on the 6 real-harm categories. Zero hard refusals across all 156 real-harm HB prompts.

MMLU 14,042 — full per-subject comparison

Overall: base 83.54% → this model 80.39% (Δ -3.15pp, well within a 3-point knowledge-preservation target).

Subjects that gained accuracy

Subjects that lost accuracy

Concentrations in moral_scenarios (-14.3pp), abstract_algebra, and public_relations are expected side-effects of removing refusal reflexes trained-in on those specific topic surfaces. Cyber (computer_security) shows a small loss vs. the small-sample uptick reported elsewhere — full 14k averages out.

Subjects unchanged (±1pp)

business_ethics, college_biology, college_mathematics, college_medicine, conceptual_physics, elementary_mathematics, international_law, jurisprudence, logical_fallacies, medical_genetics, nutrition, professional_law, sociology, virology, world_religions, astronomy, clinical_knowledge, econometrics, formal_logic, high_school_psychology, high_school_us_history, high_school_world_history, human_sexuality, philosophy, security_studies, professional_medicine, professional_psychology, high_school_government_and_politics, high_school_statistics, high_school_microeconomics.

Vision — full runtime verification

Vision tower is untouched. On a probe image (240×180 JPEG, dark-blue background with a yellow rectangular border containing the text "TEST 42"), the model returns:

"This image displays a simple, minimalist design with a solid dark blue background. Centered within the frame is a bright yellow rectangular border. Inside this border, the text 'TEST 42' appears in white, sans-serif font, centered both horizontally and vertically."

Color ✅, border color ✅, text content ✅, font style ✅, layout ✅. Full image+video+text pipeline preserved from base.

Multi-Token Prediction (MTP)

MTP head weights are untouched from the base checkpoint. mtp_num_hidden_layers = 1 remains active. Deploy with vLLM speculative-decoding to use it:

vllm serve dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8 \
  --tensor-parallel-size 2 \
  --speculative-config '{"method":"mtp","num_speculative_tokens":1}' \
  --max-model-len 262144

Reasoning modes

Same reasoning-effort levels as the Qwen3.8 base:

Compliance in off mode is highest for direct instructional requests (94.2% on real-harm HB); thinking modes at low and xhigh add reasoning depth and produce more thorough answers, at the cost of some latency and occasional additional deliberation.

Usage

Production serve command (recommended)

Use the reasoning-parser flag — without it, prior <think> blocks stay inline as raw text in subsequent multi-turn prompts, which can cause the base upstream chat template to re-emit them into new turns' prompts. With the parser, <think> content is routed to the OpenAI reasoning_content field and properly stripped from subsequent-turn prompts.

vllm serve dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8 \
  --tensor-parallel-size 2 \
  --max-num-seqs 128 \
  --reasoning-parser qwen3 \
  --tool-call-parser hermes \
  --enable-auto-tool-choice \
  --enable-prefix-caching \
  --speculative-config '{"method":"mtp","num_speculative_tokens":1}' \
  --max-model-len 262144 \
  --served-model-name qwen-flash-next

Key flags:

• --reasoning-parser qwen3 — extracts <think>...</think> into the OpenAI reasoning_content field so it doesn't leak into subsequent-turn prompts.
• --tool-call-parser hermes — parses tool calls (Qwen3.8 uses Hermes-style tool format).
• --speculative-config mtp — enables Multi-Token Prediction head for ~1.5-2× decode throughput.
• --enable-prefix-caching — reuses KV cache across identical prefixes.

Minimal serve (no MTP, no tool-calling)

vllm serve dealignai/Qwen3.8-Flash-Next-ABLITERATED-FP8 \
  --tensor-parallel-size 2 \
  --reasoning-parser qwen3 \
  --enable-prefix-caching \
  --max-model-len 262144 \
  --served-model-name qwen-flash-next

OpenAI-compatible request (Python)

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

resp = client.chat.completions.create(
    model="qwen-flash-next",
    messages=[{"role": "user", "content": "Explain the mechanism of X..."}],
    max_tokens=2000,
    temperature=0.7,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
print(resp.choices[0].message.content)

Vision request (Python)

import base64
img_b64 = base64.b64encode(open("input.jpg", "rb").read()).decode()
resp = client.chat.completions.create(
    model="qwen-flash-next",
    messages=[{"role":"user","content":[
        {"type":"image_url","image_url":{"url":f"data:image/jpeg;base64,{img_b64}"}},
        {"type":"text","text":"Describe this image."},
    ]}],
    max_tokens=500,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)

Architecture (unchanged from base)

• 48 hybrid layers (12 full-attention + 36 GDN linear-attention)
• 2,560 hidden dim, 512 experts / 10-per-token MoE, single shared expert
• FP8 quantization on expert weights, bf16 on attention/shared paths
• 262,144 native context (extendable via YaRN to 1M+)
• Multi-Token Prediction head
• Vision-language capable (image + video + text)

License

Apache-2.0 (inherited from Qwen3.8 base).

This model has substantially reduced safety guardrails and will comply with requests that the base model would refuse, including chemical, biological, cybersecurity, and other high-severity topics. Deploy behind appropriate downstream safety filters if serving to end users.