Qwen3.8-Flash-Next

2026-08-26

Date the weights (or, if earlier, the technical report) were first made public, YYYY-MM-DD, as dated by the publisher. A Hugging Face repository creation date is used only when no dated publisher page exists, with a note.

other (qwen-community-1.0)

License of the released weights as declared by the publisher (model card metadata or license file), e.g. apache-2.0, mit, or the name of a custom license.

text + image + video

Inferred from the source; see the explanation below.

Input modalities of the released checkpoint as the publisher states them, from: text, image, video, audio, joined with ' + ' in that order ("text", "text + image + video").

125B parameters

Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.

6B parameters

Parameters used per token in the forward pass, as reported by the publisher.

Not disclosed

Active parameters excluding the input embedding and the output head (unembedding). Stated by the publisher, or computed from params.active by subtracting vocab × d_model for each of the embedding and output head that params.active is stated to include (basis inferred, note with the arithmetic). Not disclosed when the scope of params.active is not stated.

125.7B parameters

Calculated: embedding + output head + all experts + other weights, Qwen/Qwen3.8-Flash-Next@de4b8e4 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Parameters of the text backbone counted from the released checkpoint's safetensors headers (embedding, output head, all experts and other weights; excluding MTP modules, vision/audio towers and n-gram lookup tables). Shown only when the counts reproduce the publisher's stated total and active parameters within 5% under some scope.

1.271B parameters

Calculated: input embedding + output head, Qwen/Qwen3.8-Flash-Next@de4b8e4 safetensors headers

Input embedding plus output head (once if tied), counted from the checkpoint.

51.2B parameters

Calculated: n-gram / per-layer embedding tables, Qwen/Qwen3.8-Flash-Next@de4b8e4 safetensors headers

Parameters in n-gram / Engram / per-layer embedding tables counted from the checkpoint. Like the token embedding, they are lookups and are excluded from params.counted_total and params.counted_active_non_embedding.

124.5B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, Qwen/Qwen3.8-Flash-Next@de4b8e4 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Language backbone weights excluding input embeddings, output head, lookup/Engram tables, MTP modules and vision/audio towers. Equals params.counted_total minus params.counted_embedding. Same validation as params.counted_total.

2.627B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, Qwen/Qwen3.8-Flash-Next@de4b8e4 safetensors headers

Largest group of language-backbone weights sharing a full layers.<index> tensor path, including every routed expert. Excludes lookup/Engram tables, embeddings, output head, MTP and other modalities. Hybrid checkpoints may store mixers and FFNs as separate layers. All-gather volume also depends on wrapping, sharding and dtype.

6.036B parameters

Calculated: non-expert weights + routed experts × 10/512, excluding embedding and output head, Qwen/Qwen3.8-Flash-Next@de4b8e4 safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Active parameters per token excluding the input embedding and output head, counted from the checkpoint as all non-expert weights plus routed-expert weights × top_k / experts. Same validation as params.counted_total.

Language-model count; excludes the 51B n-gram embedding tables and the 4B MTP module.

What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).

48

Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.

2,560

Residual stream width.

248,320

Embedding table rows (vocab_size in config, may include padding).

No

Input embedding and output head share weights.

Hashed bigram/trigram embedding at the input of layer 2, 51B params held in host memory: 16 hash heads (8 per order) of ~20M rows × 160 dims; residual-stream-gated value plus dilated short conv added to all 4 residual branches

Hashed n-gram or conditional-memory tables outside the dense backbone (Engram, n-gram embeddings), with size and placement.

53.3333

Calculated: d_model / layers

Width-to-depth ratio, d_model divided by the number of blocks.

Gated DeltaNet, QSA

Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).

3 Gated DeltaNet : 1 QSA

Inferred from the source; see the explanation below.

Token-mixer order, summarized as a pattern or layer positions. The fact's optional sequence lists each mixer in order; a ratio alone does not establish the order. Auxiliary prediction modules and standalone FFNs are excluded.

12

Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.

3

Calculated: attn.linear_layers / attn.global_layers

attn.linear_layers divided by attn.global_layers (3 for '3 Gated DeltaNet : 1 GQA').

36

Layers with a fixed-size recurrent state (linear attention, DeltaNet variants, Mamba).

24

Query heads in the global-attention layers.

2

Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.

6,144

Calculated: heads × head_dim

attn.heads × attn.head_dim, the query projection width of global layers.

12

Calculated: heads / kv_heads

attn.heads divided by attn.kv_heads in global layers.

256

Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).

partial RoPE

Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.

10,000,000

RoPE base frequency of the global layers at release.

0.25

Fraction of each head's QK dimensions that are rotated.

Yes

Queries and keys are normalized per head before the dot product.

sigmoid, elementwise on the attention output before o_proj; gate projected jointly with Q

Gate applied to the attention output before the output projection, with its placement.

QSA: MQA indexer (4 query heads, 1 key head) scores 4-token average-pooled key blocks; top-512 blocks (2048-token budget) per query plus the trailing partial block

Token-selection mechanism inside global layers, e.g. DSA lightning indexer with top-k.

No

Query/key/value projections have bias terms.

Gated DeltaNet: 16 QK heads, 48 V heads, head dim 128, short conv 4, sigmoid output gate after RMSNorm

Configuration of the linear / SSM layers (e.g. Gated DeltaNet heads and dims, Mamba-2 state size), as one string.

48

Value heads (or SSM heads) in the linear / SSM layers.

128

Head dimension of the linear / SSM layers (value head dim for DeltaNet variants, head_dim for Mamba-2).

4

Kernel size of the causal short convolution in linear / SSM layers or attention projections.

SwiGLU

FFN activation (SwiGLU, GeGLU, ...).

none

Clamping of GLU gate / up activations (e.g. SwiGLU limit 7), with the layers it applies to.

No

FFN projections have bias terms.

512

Routed experts per MoE layer.

10

Routed experts selected per token.

1

Always-on experts per MoE layer.

640

Intermediate width of one routed expert.

640

Total intermediate width of the shared expert(s).

0

Number of initial blocks that use a dense FFN instead of MoE.

softmax

Gating function over expert logits (softmax, sigmoid).

global-batch load-balancing loss

Load-balancing method, with coefficients where disclosed.

Yes

Selected expert weights are renormalized to sum to 1.

0.25

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

51.2

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.048

Calculated: params.active / params.total

params.active divided by params.total.

Not disclosed

Additional future tokens predicted during pretraining (0 when the model was trained without MTP).

1

MTP modules with distinct weights. 1 with mtp.depth > 1 means one module is applied recursively.

one decoder layer (QSA + MoE, each through Gated Residual); RMSNormed hidden state and RMSNormed next-token embedding projected by separate fc_hidden and fc_embedding

Contents of one MTP module (mixer and FFN type, norms, how the previous hidden state and next-token embedding are combined).

MoE

FFN in the MTP module: 'dense' or 'MoE', with its width when it differs from the backbone.

Yes

MTP modules use the main model's input embedding and output head.

4B parameters

Parameters of all MTP modules as reported by the publisher.

Not disclosed

Weight of the MTP loss, including its schedule.

Yes

The released checkpoint contains the MTP module weights.

RMSNorm (1+w gain)

Normalization layer type in the blocks.

pre

Pre-norm, post-norm, or sandwich (pre and post).

1e-06

Epsilon of the block normalization layers.

Gated Residual: 4 residual branches, each RMSNormed; elementwise sigmoid read gate (rank-320 bottleneck) averages branches into the block input; output written to every branch with per-branch scalar 2·sigmoid; no branch mixing

Residual design when it is not a plain sum, e.g. hyper-connections with n streams, gated residuals, depth-scaled residuals, FP32 residual accumulation.

Not disclosed

Sequence length of the main pretraining stage.

262.1K tokens

Longest sequence length trained on after context extension.

1M tokens

Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.

Not disclosed

Tokens seen in pretraining including annealing and context extension, as reported.

Muon + AdamW

Optimizer used for the bulk of parameters.

Not disclosed

Peak learning rate of the main pretraining stage.

Not disclosed

Learning rate at the end of the decay.

Not disclosed

Learning-rate schedule shape with phase boundaries.

Not disclosed

Warmup length in steps or tokens.

constant (no batch-size warmup); size not disclosed

Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.

Not disclosed

Decoupled weight decay coefficient.

Not disclosed

(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.

Not disclosed

Global gradient-norm clipping threshold.

Not disclosed

Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.

MoE router parameters normalized at initialization

Weight initialization (std or scheme) as stated by the publisher; config initializer_range alone is not evidence.

no QK-clip or SwiGLU-clip

Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.

~1/9 the training FLOPs of the 397B-A17B predecessor (~1/3 its tokens); absolute figure not disclosed

Pretraining compute as stated (GPU-hours or FLOPs).

• QSA dense distillation (indexer only)

  Tokens

  ~2B

  Seq len

  262144

  Lr

  0.001

  Batch sequences

  8

  Steps

  1000

  Source: paper · p.6–7 §2.1.2 Qwen Sparse Attention; Training Details

  only the indexer is trained for 1,000 steps, using a learning rate of 1 × 10−3 . Each step comprises 8 sequences of 256K tokens, amounting to approximately 2B training tokens in total.

  Only the final 256K continued-pretraining stages are disclosed; earlier stages and the total token count are not.

• QSA sparse training

  Tokens

  ~200B

  Seq len

  262144

  Lr

  2.5e-05

  Batch sequences

  96

  Steps

  8000

  Source: paper · p.6–7 §2.1.2 Qwen Sparse Attention; Training Details

  the backbone and indexer are jointly trained for 8,000 steps using a learning rate of 2.5 × 10−5 . Each step comprises 96 sequences of 256K tokens, totaling roughly 200B training tokens.

  Only the final 256K continued-pretraining stages are disclosed; earlier stages and the total token count are not.

Ordered pretraining stages. Each item: name, tokens, seq_len, lr, batch_tokens, batch_sequences, steps, note, quote, and source when it differs from the fact's source.

LR and batch from a hyperparameter scaling law refitted for the new architecture and Muon; target batch from step 0 (no batch-size warmup)

How peak LR, batch size or other hyperparameters were chosen for the released run (fitted scaling law with its formula, μP transfer from a proxy width, sweep), as stated.

28-layer 25B-A3B MoE (400B tokens at 4K + 80B at 32K) for architecture ablations; 35B-A3B for QSA; HP law checked on 20-layer 10.8B-A0.89B (4T tokens) and 48-layer 156B-A7B (419B tokens)

Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.

24,576 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (1024 × 16/8) B/entry

Retained KV cache divided by context length, including bounded SWA caches and indexer keys. Exports use uniform BF16 at 131,072 tokens. The website can select uniform payload precision (no quantization overhead) or sourced component formats (including scales and documented padding). This is cache payload, not peak prefill memory or persistent prefix storage.

24,576 bytes

Calculated: 24576 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (1024 × 16/8) B/entry

Long-context global KV bytes per token, including indexer keys; excludes bounded SWA caches. Compression rates and cross-layer cache reuse are included. Exports use uniform BF16; the website's storage selector also applies here. This is the field to compare with global-KV figures in reports, using the matching storage recipe.

4,096 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (1024 × 16/8) B/entry

kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).

536,870,912 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (1024 × 16/8) B/entry

kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).

3,221,225,472 bytes

Calculated: 24576 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 12 × (1024 × 16/8) B/entry

Retained KV cache for one sequence at the selected context, including SWA windows and indexer keys. Exports use uniform BF16 at 131,072 tokens. Excludes recurrent state, compressor tail state, draft/MTP caches, temporary prefill tensors, allocator metadata and persistent prefix snapshots. Does not estimate peak prefill or total serving memory.

40

Artificial Analysis Intelligence Index v4.3.2 of the post-trained model, from a dated Wayback snapshot of its Artificial Analysis page. The note names the variant (reasoning setting).