One-line: ⚠ VL — Qwen’s experimental hybrid-MoE “Flash” agentic coder (6B active / 125B total) with strong agentic-coding evals, caught by the per-org sweep (not the text-gen streams; it’s image-text-to-text). ⚠ All scores below are lab-reported (Qwen’s own card) — no independent source yet on this 3-day-old drop. Custom architecture Qwen4ExpForConditionalGeneration requires trust_remote_code.
Coding benchmarks
Benchmark
Score
Source
HumanEval
—
—
HumanEval+
—
—
MBPP
—
—
LiveCodeBench
91.9 (v6)
lab-reported
Aider polyglot
—
—
SWE-bench Verified
— (publishes SWE-bench Pro 62.5 instead)
lab-reported
SWE-bench Multilingual
81.0
lab-reported
DeepSWE 1.1 (agentic stand-in)
58.7
lab-reported
Agentic / tool-use benchmarks
Benchmark
Score
Source
BFCL
— (publishes Toolathlon Verified 73.5 instead)
lab-reported
τ-Bench
—
—
ToolACE
—
—
GAIA
—
—
Toolathlon Verified (Pass@1, stand-in)
73.5
lab-reported
Agents’ Last Exam (frontier agentic)
Pass@1 24.3 / Score 51.2
lab-reported
CoWorkBench (long-horizon office)
73.9
lab-reported
Reasoning (context only)
Benchmark
Score
Source
MMLU / MMLU-Pro
—
—
GPQA Diamond
91.7
lab-reported
MATH
—
—
AIME 2025
—
—
HLE
35.9
lab-reported
IFBench (instruction following)
81.3
lab-reported
Efficiency
Active params: 6B active (125B total, +51B n-gram embedding, +4B MTP)
Context length: 262,144 natively, extensible to 1,000,000 tokens
VRAM (fp16 rough): ~250GB at fp16 (125B total) → ⚠ large; must be quantized for local/agent use. 6B active keeps inference compute low but the checkpoint download is huge.
Qwen3.8-Flash-Next is an experimental “Flash” variant that pairs a 6B-active MoE body with Gated DeltaNet + Qwen Sparse Attention (QSA) hybrid attention, targeted squarely at agentic workloads — QSA cuts long-context latency, preserved-thinking keeps reasoning traces for multi-turn agents, and it natively runs 262K context (1M with RoPE/YaRN scaling). Its agentic-coding numbers are legitimately strong for a 6B-active model: SWE-bench Pro 62.5, SWE-bench Multilingual 81.0, DeepSWE 1.1 58.7, LiveCodeBench v6 91.9, Toolathlon 73.5. It edges its own larger sibling Qwen3.8-27B and DeepSeek-V4-Flash-0731 on most agentic rows. Caveats: it’s a VLM (image-text-to-text), all numbers are lab-reported with no independent confirmation yet, and the custom architecture means local deployment requires repo model code and heavy quantization (~250GB fp16). For this user’s “runs locally” bar, the 6B-active MoE is attractive but the total-param download and custom-arch requirement are real friction right now.