Solar Open 2

Upstage (Upstage Solar Team) · 2026-07-22 · analysed 2026-08-10 · skill v0.4.0

sparse-MoE hybrid text-only open-weights frontier frontier-generalist

Design overview

Solar Open 2 is a 250B-parameter sparse MoE (15B active) that scales the Solar Open 1 backbone — 48 layers, hidden 4,096, head dim 128, 64 query / 8 KV heads, 196,608-token vocab — while replacing its all-softmax GQA stack with a 3:1 linear:softmax hybrid. Every block of four layers runs the softmax layer first (S-L-L-L, opposite to Kimi Linear and Qwen3.5), then three KDA linear-attention layers, each paired with an MoE FFN (320 routed experts, top-8, one shared expert, SiLU, per-expert width 1,280). Three extensions define the design: NoPE — no positional encoding anywhere, so the linear layers' sequential state carries all token order and the model extrapolates in principle without bound; a sigmoid output gate on the softmax layers that adds query-dependent sparsity and suppresses attention-sink collapse; and the negative-eigenvalue extension of the gated delta rule (β = 2σ(·) ∈ (0,2)), which widens the state-transition eigenvalues to [−1,1] so the recurrent state can erase and invert, not merely decay — restoring provable state-tracking (parity, modular counting) that [0,1]-eigenvalue cores lack. The 1M context is reached by staged training: 10T general pre-training, 1T intensive, then 0.9T length expansion, with four checkpoints merged. Training efficiency is attacked from two sides: a selective weight transfer that reuses only the 5.69B-parameter shape-compatible skeleton from Solar Open 1 (no distillation), and a value-per-token data recipe (20T pool → 10T, 4:6 real:synthetic). Post-training runs SFT → multi-domain RL → twelve domain specialists → Multi-teacher On-Policy Distillation, folding the specialists into one model with exact full-vocabulary KL on the student's own rollouts.

What the report claims is novel

Atlas assessment: The block is a well-executed synthesis of mechanisms already in the atlas: KDA linear attention (moonshot-kimi-k3), GQA (llama-3.1 lineage), NoPE hybrid stacking (kimi-k3), and sigmoid output gating (Qiu et al. 2025, per the report's own citation). The two genuinely differentiating choices are (1) softmax-first (S-L-L-L) ordering, explicitly opposite to Kimi Linear/Qwen3.5's L-L-L-S, and (2) the negative-eigenvalue extension (β = 2σ(·) ∈ (0,2)) that widens the gated delta rule's state-transition eigenvalues to [−1,1] — an incremental but real generalization of the KDA recurrence that Kimi's own lower-bounded-decay parameterization does not provide. At the system level, selective weight transfer across an architectural change without distillation (only 2.3% of parameters reused) and exact full-vocabulary MOPD at 250B scale are notable engineering contributions. The 250B-A15B footprint with 1M context and 3:1 linear:softmax ratio is essentially the kimi-k3 hybrid recipe at one-ninth the total size with GQA instead of MLA.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilySolar Open 2[1]
OrganisationUpstage (Upstage Solar Team)[2]
Release date2026-07-22[3]
Report URLhttps://huggingface.co/upstage/Solar-Open2-250B/raw/main/Solar_Open_2_Tech_Report.pdf[4]
Licensenot disclosed
Open weightsyes[5]

Scale

Total parameters250B[6]
Active parameters / token15B[7]
Layers48[8]
Hidden dimension4096[9]
FFN inner dimension1280[10]
FFN ratio0.3125
Query heads64[11]
KV heads8[12]
Head dimension128[13]
Vocabulary size196608[14]
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention varianthybrid[15]
Attention layer patternper block of four: one softmax (GQA) layer first, then three linear (KDA) layers — S-L-L-L; repeated 12 times; 12 softmax (25%) + 36 linear (75%) layers; softmax-first ordering explicitly contrasts with the L-L-L-S ordering of Kimi Linear and Qwen3.5[16]
Depth mixingsequential-residual
ActivationSiLU[17]
Expert count320[18]
Experts per token8[19]
Shared experts1[20]
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
MethodNoPE[23]
RoPE base frequencyn/a[24]
Partial RoPEn/a[25]
Typenot disclosed
Placementpre[26]
QK-normnot disclosed
Attention sinksno
Softcappingno

Context

Trained context length1,048,576[27]
Deployed context length1,048,576[28]
Extension methodnone — no positional-encoding extension needed (NoPE); the 1M window is reached by staged training, with Stage 4 (Length Expansion, 0.9T tokens) extending the context window to 1M[29]

Tokenizer

Algorithmcustom byte-level BPE[30]
Notesinherited unchanged from Solar Open 1; trained on a corpus oversampling Korean and target domains, with digit splitting and whitespace preservation for arithmetic and code fidelity; 4.41 bytes/token on Ko-GDPval task prompts (first among 12 tokenizers compared, ~24% above the best global model MiniMax-M3); identical vocabulary across generations is a precondition for transferring embedding/output-layer weights[31]

Training

Training tokens12T[32]
Data composition20T initially cleaned pool refined to 10T: every dataset receives an in-house quality score and rarity tracking by source/type; exact and semantic deduplication; mixture-ratio optimization via ablations — real-to-synthetic ratio 4:6, math and code each at least 15%, English share above 80%[33]
Curriculum / stagingfour-stage pre-training: Stage 1 Selective Weight Transfer (2.3% of parameters from Solar Open 1), Stage 2 General Pre-training (10T tokens, broad-coverage mixture), Stage 3 Intensive Pre-training (1T tokens, higher quality threshold on the retained subset), Stage 4 Length Expansion (0.9T tokens, long-document data incl. repo-level code, extends context to 1M); four intermediate checkpoints merged into the final model (highest PLM score 0.745)[34]
Optimizernot disclosed
LR schedulenot disclosed
Batch schedulenot disclosed
Precisionnot disclosed
ParallelismContext Parallelism (CP): student and teacher models are sharded by CP rank during RL/MOPD training, and teacher transport ships only the matching CP shard, skipping the CP all-gather; other parallelism dimensions (TP/PP/EP) not disclosed[35]
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[36]
Preference optimisationother[37]
Reasoning trainingMulti-domain RL strengthens core reasoning with verifiable rewards over a broad set of STEM-centered tasks; twelve domain specialists (reasoning: math/STEM/code; agents & tools: coding, general agents, officework) trained via domain SFT and domain-specific RL with verifiable rewards; SFT covers the full range of reasoning and agent tasks[38]
DistillationMulti-Teacher On-Policy Distillation (MOPD): the student is trained on its own rollouts against twelve domain-specialist teachers; per-position distillation KL computed in closed form over the full 196,608-token vocabulary (teacher ships pre-lm_head hidden states, student rebuilds logits in 1,024-token tiles); each sample is routed to exactly one teacher, keeping teacher compute O(batch); note the selective weight transfer from Solar Open 1 is initialization, explicitly 'without distillation'[39]

Modality

Typetext-only
Attachmentn/a[40]

Inference efficiency

KV-cache designhybrid cache: 12 softmax GQA layers keep a KV cache with 8 KV heads, while 36 linear KDA layers keep a fixed-size recurrent state and no KV cache — total cache constant in sequence length, roughly one quarter of an all-softmax stack's memory
Quantisation shippednone disclosed
Speculative decodingnone disclosed
Serving optimisationsnone disclosed

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU-Pro86.2Table 3
GPQA-Diamond86.3Table 3
HLE (w/o tools)28.8Table 3
LiveCodeBench (v6)92.4Table 3
ArtifactsBench55.9Table 3
HMMT260293.9Table 3
AIME202695.7Table 3
Multi-Challenge61.0Table 3
IFBench80.0Table 3
AA-LCR62.3Table 3
SWE-Bench Verified70.4Table 3
Terminal Bench Hard28.3Table 3
APEX-Agents16.6Table 3
MCP-Atlas58.2Table 3
τ3 (banking)19.6Table 3
GDPval-AA v2 (ELO)1128Table 3
KMMLU-Pro78.4Table 4
CLIcK90.7Table 4
Ko-AIME'2597.7Table 4
HRM8K92.2Table 4
KorMedMCQA93.0Table 4
Ko-GDPval86.8Table 4
Korean suite Avg.85.4Table 4

Provenance references

  1. stated · Title ('Solar Open 2 Technical Report')
  2. stated · Title page byline ('Upstage Solar Team'); website upstage.ai
  3. stated · Title page ('Date: Jul 22, 2026'); arXiv:2607.20062v1 [cs.CL] 22 Jul 2026
  4. stated · pipeline
  5. stated · §1 ('an openly available, nationally grounded frontier model'); §4 (MOPD consolidates twelve teachers 'into the single released model'); Acknowledgments ('the development and release of Solar Open 2')
  6. stated · Abstract, §2, Table 1 ('250B total parameters')
  7. stated · Abstract, §2, Table 1 ('15B active parameters per token')
  8. stated · §2, Table 1 ('Num Layers 48')
  9. stated · §2, Table 1 ('Hidden Size 4,096')
  10. stated · Table 1 ('MoE Intermediate Size 1,280' — per-expert FFN inner dimension)
  11. stated · §2, Table 1 ('64 query and 8 key-value heads')
  12. stated · Table 1 ('Attention Heads (Q / KV) 64 / 8'), note a: 'KV heads apply only to the softmax layers; the linear layers use 64 heads and keep no KV cache'
  13. stated · §2, Table 1 ('Head Dimension 128')
  14. stated · §2, Table 1 ('a 196,608-token vocabulary')
  15. stated · §2.2, Table 1 ('Hybrid: [Softmax ×1, Linear ×3] ×12'; GQA softmax layers + KDA linear layers)
  16. stated · §2.2 ('interleaves one softmax-attention layer with three linear-attention layers in every block of four... so that 12 of the 48 layers (25%) are softmax and 36 (75%) are linear'; 'the softmax layer comes first (S-L-L-L)')
  17. stated · Table 1 ('Activation: SiLU')
  18. stated · §2, Table 1 ('Num Experts (Routed) 320')
  19. stated · Table 1 ('Experts per Token (Top-k) 8')
  20. stated · §2, Table 1 ('Num Shared Experts 1'; 'a single shared expert with no dense layers')
  21. n/a · §2.2 (no rescaling factor applies; no RoPE)
  22. stated · Table 1 ('Context Length 1,048,576 (1M)')
  23. stated · §2.2 ('no positional encoding anywhere in the model'), Table 1 ('Positional Embedding: NoPE')
  24. n/a · §2.2 (no RoPE anywhere in the model; Solar Open 1's RoPE θ = 10⁶ is listed only as the discarded predecessor setting)
  25. n/a · §2.2 (no RoPE)
  26. stated · §2 ('pre-norm residuals')
  27. stated · §3.3 (Stage 4 Length Expansion 'extends training to a 1M context length'), Table 1 ('Context Length 1,048,576 (1M)')
  28. stated · Table 1 ('Context Length 1,048,576 (1M)'); §2.2 ('usable context window beyond 1M tokens')
  29. stated · §2.2, §3.3
  30. stated · §2.1 ('a custom byte-level BPE tokenizer with a 196,608-token vocabulary')
  31. stated · §2.1
  32. stated · §3, Figure 5 caption ('Stages 2–4 then train on about 12T tokens in total: General Pre-training (10T), Intensive Pre-training (1T), and Length Expansion (0.9T)')
  33. stated · §3.2
  34. stated · §3, §3.3
  35. stated · §4.2.2 ('Under context parallelism, the teacher's CP rank r already holds exactly the shard student CP rank r needs')
  36. stated · §4 ('Post-training of Solar Open 2 proceeds in four stages: (1) Supervised Fine-Tuning (SFT)...'; SFT 'targets the full range of reasoning and agent tasks')
  37. stated · §4.2.1 (fully asynchronous RL with verifiable rewards; 'Retaining token-level ratios as in GRPO (Shao et al., 2024)' with trust-region clipping, stale-trajectory masking via fresh-token ratio threshold ρ, and sample dropping — GRPO-style but modified, so tagged other)
  38. stated · §4, §4.2.2
  39. stated · §4.2.2, §3.1
  40. n/a · §1