ZAYA1

Zyphra · 2026-05-06 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE efficient-attention text-only undisclosed medium reasoning-specialised

Design overview

ZAYA1-8B is Zyphra's reasoning-focused MoE: 8.4B total parameters, 0.76B active, 40 layers, 16 experts per layer with top-1 routing and no shared or residual experts. Its block makes three departures from standard MoE transformers. First, sequence mixing uses Compressed Convolutional Attention in GQA mode (CCGQA): attention is performed in a compressed latent space after low-rank down-projections and lightweight sequence-mixing convolutions (a short convolution plus grouped head-wise convolution) precondition queries and keys; half the value heads are time-delayed by one token; query/key mean skip connections and an RMSNorm that applies a learned head-wise temperature to the keys (parameterized linearly as T, not exp(T)) keep attention logits bounded under QK-norm. With 8 query heads sharing 2 KV heads on top of 2× query compression, the KV cache is 8× smaller than full MHA. Second, the ZAYA1 router replaces the linear router with a down-projection to 256 dims, an Exponential Depth Averaging blend with the previous layer's routing representation (a DenseFormer-style depth-weighted averaging variant), and a three-layer GeLU MLP, with PID-controller-style bias balancing (AdamW on p_e − 1/E) rather than DeepSeek-style sign updates. Third, learned residual scaling (per-channel α, β on both the residual stream and layer outputs) controls residual-norm growth through depth. The model was trained from scratch for reasoning — long-CoT data in every pretraining phase with answer-preserving trimming, a 32K midtraining and 131K SFT on AMD MI300X + Pollara hardware, then a four-stage verifiable-RL cascade — and is co-trained with Markovian RSA, a test-time-compute workflow that recursively aggregates candidate reasoning traces while carrying forward only bounded tails (e.g. 4K of a 40K budget), lifting AIME'25 to 91.9%.

What the report claims is novel

Atlas assessment: The three architectural changes are refinements of known families rather than new mechanisms: CCA/CCGQA is softmax attention moved into a compressed latent space with convolutional preconditioning (a compute-efficiency variant of GQA — the atlas's first compressed-latent attention, but conceptually adjacent to MLA's latent KV); the ZAYA1 router upgrades DeepSeek-V3-style bias balancing (PID update via AdamW) and adds a DenseFormer-style depth-averaged MLP router; residual scaling is a cheap alternative to Qwen-style attention gating. Top-1 routing with no shared experts is an aggressive sparsity choice not seen in atlas peers (deepseek-v3/mixtral use top-8 with shared experts). The genuinely novel contribution is Markovian RSA with training-time integration — the first TTC workflow trained into the model in the atlas — though it is an inference/training scheme, not an architecture change. vs deepseek-r1: same reasoning-specialised, verifiable-RL recipe at far smaller scale. The report is exemplary on training disclosure (phases, mixtures, precision, optimizer, stability engineering) but omits the expert FFN activation, tokenizer algorithm, training compute, and any weights-release statement.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyZAYA1[1]
OrganisationZyphra[2]
Release date2026-05-06[3]
Report URLhttps://arxiv.org/abs/2605.05365[4]
Licensenot disclosed
Open weightsnot disclosed

Scale

Total parameters8.4B[5]
Active parameters / token760M[6]
Layers40[7]
Hidden dimension2048[8]
FFN inner dimension4096[9]
FFN ratio2.0
Query heads8[10]
KV heads2[11]
Head dimension128[12]
Vocabulary size262272[13]
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention variantGQA[14]
Attention layer patternuniform — all 40 layers use CCGQA (compressed latent-space attention); no full-attention or SSM layers anywhere in the stack[15]
Depth mixingsequential-residual
Activationnot disclosed
Expert count16[16]
Experts per token1[17]
Shared experts0[18]
RoutingZAYA1 router: down-projection to 256 dims, Exponential Depth Averaging blend with the previous layer's routing representation, then softmax(MLP(RMSNorm(r))) scores plus learned bias b; top-1 selection via argmax_e(s_e + b_e)[19]
Load balancingAuxiliary-loss-free PID-controller-style bias balancing: bias gradient ∇b_e = p_e − 1/E (empirical routing fraction minus uniform), updated with AdamW across the global batch; builds on DeepSeek-V3's bias scheme[20]
Expert granularityRelatively fine-grained: 16 experts per layer with a 2× hidden expansion (expert FFN 4096 pre-activation / 2048 post-activation)[21]
MethodRoPE[25]
RoPE base frequency5000000[26]
Partial RoPEyes[27]
TypeRMSNorm[28]
Placementpre[29]
QK-normyes[30]
Attention sinksno
Softcappingno
OtherLearned key temperature T scaled linearly (not exp(T)) to bound the maximum attention logit at T·√d_h under QK-norm (stability modification of standard CCA); Query/key mean skip connections enforcing representational similarity; Learned residual scaling (α, β per layer on residual stream and layer output) controlling residual-norm growth through depth, with α initialized to ones and β to zeros[31]

Context

Trained context length131,072[32]
Deployed context length131,072
Extension methodStaged RoPE base-frequency increase: 10K (4K context) → 1M (32K midtraining) → 5M (131K SFT); no interpolation factor stated[22]

Tokenizer

Algorithmnot disclosed
NotesGemma3 tokenizer, 262,272 vocabulary size[33]

Training

Training tokens13.86T
Data compositionBase pretraining: broad web-crawl distribution with code, math, multilingual, and reasoning data mixed in progressively; phase 2 upweights code, math, reasoning, and instruction-formatted data. 32K midtraining mixture: 86.1% long-CoT reasoning traces, 5.7% web/synthetic web/multilingual, 0.8% natively long-context, 3.0% code corpus, 3.0% math/STEM, 1.4% short instruction/few-shot. 131K SFT mixture: 75.0% long-CoT, 9.8% web, 6.4% natively long-context, 5.0% code corpus/code SFT, 2.6% math/STEM, 1.2% short instruction. Individual dataset names are omitted.[34]
Curriculum / stagingStaged: 4K base pretraining (two phases) → 32K reasoning-focused midtraining (1.2T tokens) → 131K SFT (660B tokens); long-CoT reasoning data in every pretraining/midtraining phase (majority of the midtrain mix) with stage-aware answer-preserving (AP) re-trimming at each context length; RL cascade follows post-training[35]
OptimizerMuon with AdamW RMS matching (pretraining, midtraining, SFT); RL: momentum-free Muon for matrix-valued actor weights with AdamW for embeddings and LM head[36]
LR schedulenot disclosed
Batch scheduleRL: minibatches of 128 prompts with rollout group size G=16, 131,072-token budget per GPU per microbatch with sequence packing and dynamic batching; pretraining run example (Appendix B): G=4096, s=4096, 4 bytes/token[37]
PrecisionBF16 weights and activations with a matched FP32 op set on trainer and inference engine: fused cross-entropy + LM-head matmul, CCA cache state, QK-norm, QK-mean, RMSNorm, router softmax, residual-stream additions[38]
ParallelismData parallelism plus all-gather KV context parallelism (2 ranks at 32K, 8 ranks at 131K); FSDP2 shard size 4 with sequence parallelism disabled; post-training uses asynchronous PipelineRL with disjoint rollout/trainer GPU pools (2–5× more rollout workers, weight sync every 2 trainer iterations)[39]
HardwareAMD MI300X GPUs (8× per node, Infinity Fabric intra-node) with AMD Pensando Pollara 400 networking (400 Gb/s NICs); dual-socket Intel Xeon Platinum 8570 nodes; 25.6 TB NVMe storage per compute node[40]
Disclosed computenot disclosed

Post-training

SFTyes[41]
Preference optimisationother[42]
Reasoning trainingTrained from scratch for reasoning: long-CoT data in all pretraining/midtraining phases with answer-preserving trimming; verifiable RL over math and puzzles (232-step warmup; 400-task RLVE-Gym adaptive difficulty curriculum with Thompson-sampling IRT calibration), code (CodeI/O prediction, CodeARC reconstruction, falsification synthetic environments built from competitive-programming references), and TTC aggregation prompts (Markovian RSA expert- and self-aggregation); per-rollout response budgets up to 81,920 tokens; no dedicated multi-turn agentic RL stage[43]
Distillationnone[44]

Modality

Typetext-only
Attachmentn/a[45]

Inference efficiency

KV-cache designCCGQA: KV-cache compression 8× relative to full multi-head attention (2 KV heads shared across 8 query heads on top of 2× query compression); CCA's compressed latent KV representation kept activation and KV-cache memory low enough for 131K-context SFT and RL[46]
Quantisation shippednone disclosed[47]
Speculative decodingnone disclosed[47]
Serving optimisationsnone disclosed[47]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
AIME'2689.1Table VII
HMMT'26 Feb.71.6Table VII
IMO-AnswerBench59.3Table VII
APEX-shortlist32.2Table VII
LiveCodeBench-v664.8Table VII
GPQA-Diamond71.0Table VII
MMLU-Pro74.2Table VII
IFEval85.6Table VII
IFBench52.6Table VII
EQBench73.0Table VII
Creative Writing v363.0Table VII
BFCL-v440.5Table VII
tau236.3Table VII
AIME'25 (Markovian RSA 40K/4K)91.9Table XI, §VI-C
HMMT'25 Feb. (Markovian RSA 40K/4K)89.6Table XI, §VI-C

Provenance references

  1. stated · Title, §VII ('ZAYA1-8B, the first and smallest model in the ZAYA-1 family')
  2. stated · Title page byline (Zyphra, San Francisco, CA)
  3. stated · Header: 'arXiv:2605.05365v1 [cs.AI] 06 May 2026'
  4. stated · pipeline
  5. stated · Table I (Total parameters 8.4B)
  6. stated · Table I (Active parameters 0.76B)
  7. stated · Table I (Transformer layers 40); Appendix D (40 MoE layers)
  8. stated · Table I (Hidden dimension 2048)
  9. stated · Table I (Expert FFN width: 4096 pre-activation / 2048 post-activation)
  10. stated · Table I (CCA query heads 8)
  11. stated · Table I (KV heads 2); Appendix C ('CCGQA with 2 KV heads for 8 query heads')
  12. stated · Table I (Head dimension 128)
  13. stated · Table I (Tokenizer: Gemma3 tokenizer, 262,272 vocabulary size)
  14. stated · Table I (Attention variant: CCGQA with CCA preconditioner); Appendix C
  15. stated · §II-A ('ZAYA1-8B uses Compressed Convolutional Attention (CCA) for the attention block'), Table I, Appendix C
  16. stated · Table I (Experts per MoE layer 16)
  17. stated · Table I (Routing: Top-1, no residual expert); §II-A2 (k=1)
  18. stated · §II-A3 ('we trained with top-k equal to 1 and without residual experts')
  19. stated · §II-A2, Eqs. 1–4
  20. stated · §II-A2, Eq. 5 (cites DeepSeek-AI 2025a)
  21. stated · §II-A3, Table I
  22. stated · Table II, §III
  23. n/a · Table II (no rescaling factor disclosed; extension via base-frequency increase)
  24. stated · Table II (SFT phase at 131K context)
  25. stated · §II-A3 ('We applied RoPE (Su et al., 2023) to half the channels in each head, leaving the other half without position embeddings')
  26. stated · Table II
  27. stated · Table I (Position embeddings: 50% RoPE on each head); §II-A3
  28. stated · §II-A2 Eq. 3 (router: RMSnorm(r_l)); §II-A3 Eq. 6 (Layer(RMSnorm(x_l))); Appendix C (RMSNorm head-wise temperature on keys)
  29. stated · §II-A3 Eq. 6 (x_{l+1} = Res-scale_res(x_l) + Res-scale_out(Layer(RMSnorm(x_l))))
  30. stated · Appendix C ('The scaled query-key inner product is then upper bounded (assuming QK norm) to T·d_h/√d_h = T√d_h'); §IV-D (QK-norm and QK-mean in the FP32 op set)
  31. stated · Appendix C, §II-A3 Eq. 6
  32. stated · Table II (SFT phase at 131K context, 660B tokens, RoPE base 5M)
  33. stated · Table I
  34. stated · §III, Tables II–III
  35. stated · §III, §III-A
  36. stated · §III, §IV-E
  37. stated · §IV-B1, §IV-C, Appendix B
  38. stated · §IV-D
  39. stated · §III, §IV-C, §IV-B1
  40. stated · Table I, Appendix A (Table XIV)
  41. stated · §IV-A ('SFT phase at 131K context for 660B tokens'; chat, IF, code, math, reasoning, tool-calling and TTC-aggregation mixture; best-fit decreasing bin packing)
  42. stated · §IV-B, §IV-B1, §IV-B7
  43. stated · Abstract, §III-A, §IV-B3–B6, §VI-B
  44. stated · Abstract ('ZAYA1-8B was trained from scratch for reasoning')
  45. n/a · modality.type text-only
  46. stated · Table I, §III, Appendix C
  47. stated · report