Model datasheet · Qwen3.5
Qwen3.5 397B-A17B
The largest model on this site and the top of the Qwen3.5 line. Gated DeltaNet plus sparse MoE: only 15 of its 60 layers are full attention, so 256K of context costs about 8GB of cache — remarkably cheap for its size. The weights are the problem, not the cache.
- Vendor
- Alibaba
- Architecture
- MoE · 397B total / ~17B active
- Context
- 262,144 tokens
- License
- Apache-2.0
- Released
- 2026-02-16
- Vision
- Yes
→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)
Architecture, from its own config.json
60 layers · 45 linear / recurrent · 15 full attention
Only 15 of 60 layers hold a cache at all, which is why its memory grows so much more slowly with context than its size suggests.
32 attention heads over 2 KV heads (GQA 16:1), head dim 256 · mixture of experts, 17B of 397B active per token across 10 experts · vocabulary 248,320
Where the memory goes: weights by band above, plus a KV cache of 240MB at 8K · 960MB at 32K · 3.8GB at 128K, plus the runtime's own working memory — dominated by the output projection over that vocabulary.
Inside one layer, from its own tensors
45 × linear / recurrent layer
15 × full-attention layer
The green line down the right of each half is the residual: it leaves the spine above the norm, skips the operation entirely and rejoins at the +. Where attention is drawn open, q, k and v are scaled to their real output widths — a model that shares key and value heads across many query heads shows one wide box beside two narrow ones.
Around the stack: an embedding of 248,320 × 4,096, a final norm, and a separate output projection.
On the norm names: despite what it is called, post_attention_layernorm is drawn after the residual add here, which is where this family applies it — it normalises the input to the feed-forward, not the output of attention. Models that carry a separate pre_feedforward_layernorm use it the other way, and are drawn that way.
§1 Characteristics by quantization band
Q2–Q3 — 2–3 bit
177GB weights · e.g. IQ2_M , Q3_K_M + KV: 240MB at 8K · 960MB at 32K · 3.8GB at 128Kdata checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q3_K_M · aggregated · 2026-08-31
Editorial derivation: two points down. At 3-bit a MoE this large loses expert-routing precision first, which surfaces as dropped tool arguments and broken multi-file edits before prose visibly suffers.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q3_K_M · aggregated · 2026-08-31
Editorial derivation: two points down. At 3-bit a MoE this large loses expert-routing precision first, which surfaces as dropped tool arguments and broken multi-file edits before prose visibly suffers.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-08-31
Editorial derivation: one point down from the measured bf16 figure. 4-bit is the band a model this size is actually run at, and the loss is real but modest.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-08-31
Editorial derivation: one point down from the measured bf16 figure. 4-bit is the band a model this size is actually run at, and the loss is real but modest.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-08-31
Editorial derivation: one point down from the measured bf16 figure. 4-bit is the band a model this size is actually run at, and the loss is real but modest.
data checked Oct 2026
- quant-degradation-community — editorial band derivation @ Q4_K_M · aggregated · 2026-10-04
Editorial derivation: carried from the EQ-Bench-anchored full band; creative writing is not quant-sensitive.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Aug 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-08-31
Editorial derivation: no drop from the measured bf16 figure. 6-8 bit is effectively lossless on a model this size — the differences that show up at 4-bit and below are not measurable here. Recorded explicitly rather than left to band fallback, which would otherwise borrow the 4-bit number and subtract a point.
data checked Oct 2026
- quant-degradation-community — editorial band derivation @ Q6_K · aggregated · 2026-10-04
Editorial derivation: carried from the full band; creative writing is not quant-sensitive.
FP16 — full precision (fp16/bf16)
793GB weights · e.g. bf16 + KV: 240MB at 8K · 960MB at 32K · 3.8GB at 128Kdata checked Aug 2026
- vendor-model-card — SWE-bench Verified = 76.4 @ bf16 · vendor-reported · 2026-08-31
LiveCodeBench v6 83.6, Terminal Bench 2 52.5. Beats GLM-4.7's 73.8 on SWE-bench.
data checked Aug 2026
- vendor-model-card — MMLU-Pro = 87.8 @ bf16 · vendor-reported · 2026-08-31
No hallucination benchmark published; this is a knowledge score, not a truthfulness one.
data checked Aug 2026
- vendor-model-card — IFEval = 92.6 @ bf16 · vendor-reported · 2026-08-31
IFBench 76.5. Both are program-checkable, so there is little interpretive room.
data checked Aug 2026
- vendor-model-card — AIME26 = 91.3 @ bf16 · vendor-reported · 2026-08-31
GPQA 88.4, MMLU-Pro 87.8 — the highest GPQA figure in this catalogue.
data checked Aug 2026
- vendor-model-card — BFCL-V4 = 72.9 @ bf16 · vendor-reported · 2026-08-31
Terminal Bench 2 52.5 — the best terminal-agent figure here, though BFCL trails GLM-4.7's Tau2 result.
data checked Oct 2026
- eqbench-creative — EQ-Bench Creative Writing v3 rubric score (0-20) = 16 @ bf16 · aggregated · 2026-10-04
Normalised Elo 1478.2, 68th of 140, retrieved 2026-10-04. Rubric level with Gemma 4 31B and 26B-A4B (16.01, 16.03), below Muse Glimmer 30B (16.26, scored 8). Quant not stated by EQ-Bench (presumed full-precision API serving). Independent; rubric judged by Claude Sonnet 4. EQ-Bench's normalised Elo moves as models are added, so the rubric score is the figure stored. Data file behind https://eqbench.com/creative_writing.html; the date the row was added is not published.
- eqbench-creative — EQ-Bench Creative Writing Longform score (0-100) = 68.3 @ bf16 · aggregated · 2026-10-04
Longform: the highest figure of any model on this site, ahead of Gemma 4 31B (56.5) and Qwen3.8 27B (53.3). Quant not stated by EQ-Bench (presumed full-precision API serving). Data file behind the Longform page on eqbench.com.
§2 Known issues & what fixes them
The unusual shape of this model is that context is cheap and weights are not. Three of every four layers use linear attention, so the KV cache is only 30KB per token — a full 256K context costs about 8GB, less than GLM-4.7 spends on 32K. But the 4-bit weights are 244GB. Nothing about a long context will stop you here; holding the model will.
Architectural, and mostly good news. Size the machine for weights and treat context as close to free.
Evidence · 2 sources · community-consensus
- Qwen3.5-397B config.json: layer_types, 3 linear per 1 full attention across 60 layers (vendor, 2026-08-31)
- Unsloth Qwen3.5 run guide (blog, 2026-08-31)
Every number scored here is Alibaba's own. Qwen3.5-397B is a flagship release with headline figures that beat much larger closed models, and no independent replication of those figures exists in this dataset. Treat them as claims. Its factuality and long-context cells are blank rather than estimated because nobody has published a hallucination or RULER result for it.
Nothing to work around — this is a caveat on the evidence, not on the model.
Evidence · 2 sources · community-consensus
- Qwen3.5-397B-A17B model card: all benchmark figures are vendor-reported (vendor, 2026-08-31)
- unsloth GGUF quant ladder — community distribution, no independent benchmarks (other, 2026-08-31)