Solar Open 2
Upstage · Solar · open weights
- GQA
- Linear attention
- DeltaNet / KDA
- Hybrid
- RoPE
- NoPE layers
- Pre-norm
- MoE
- Shared expert
Facts and where they come from
| Released | 2026-07 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | other | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 250B | labmodel cardREADME: Total Parameters 250B (250,287,794,944) |
| Active parameters | 15B | labmodel cardREADME: Active Parameters 15B |
| Context length | 1M tokens | labmodel cardREADME: Context Length 1M |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for solar_open2: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for solar_open2: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | no | codemodelling codeno q/k normalisation in the attention block |
| Positional encoding | RoPE; attention layers use no RoPE; the linear-attention layers carry order | config.jsonconfig.jsonuse_rope: false |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for solar_open2: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
12× GQA 64q/8kv + 36× Kimi Delta Attention. Each column is one layer: its token mixer above, its feed-forward block below. Paler columns reuse another layer’s keys and values.
Modelled costs
From the cost model, batch size 1. Totals the lab states are in the table above; differences come from rounding, from what a lab counts, or from parts the model does not describe (listed on the about page).
| Parameters (modelled) | 251B |
|---|---|
| Active per token (modelled) | 15.9B |
| Without embeddings and output head | 250B total, 14.3B active |
| Published weights (Hugging Face count) | 250B |
| KV cache per token, BF16 (layers that grow with context) | 48 KiB |
| KV cache + state at 1M tokens, BF16 | 48.1 GiB |
| Decode FLOPs per token at 4K context | 32.1 GFLOP |
| Prefill FLOPs for a 4K prompt | 122 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 4,096 | config.jsonconfig.jsonhidden_size |
| vocab | 196,608 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.full.type | attn | codemodelling codeattention block |
| mixers.full.heads | 64 | config.jsonconfig.jsonnum_attention_heads |
| mixers.full.kv_heads | 8 | config.jsonconfig.jsonnum_key_value_heads |
| mixers.full.head_dim | 128 | config.jsonconfig.jsonhead_dim |
| mixers.full.gate | elementwise | codemodelling codeuse_gqa_gate: true |
| mixers.kda.type | kda | codemodelling codeKimi Delta Attention |
| mixers.kda.k_heads | 64 | config.jsonconfig.jsonlinear_attn_config.num_heads |
| mixers.kda.v_heads | 64 | config.jsonconfig.jsonlinear_attn_config.num_heads |
| mixers.kda.k_head_dim | 128 | config.jsonconfig.jsonlinear_attn_config.head_dim |
| mixers.kda.v_head_dim | 128 | config.jsonconfig.jsonlinear_attn_config.head_dim |
| mixers.kda.conv_kernel | 4 | config.jsonconfig.jsonlinear_attn_config.short_conv_kernel_size |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 320 | config.jsonconfig.jsonn_routed_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 1,280 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| ffns.moe.shared | 1 | config.jsonconfig.jsonn_shared_experts |
| ffns.moe.d_shared | 1,280 | config.jsonconfig.jsonmoe_intermediate_size |
| layout | 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe · 1× full/moe · 3× kda/moe | config.jsonconfig.jsongqa_layers |
Sources
- config.json @ 9190fbe
- model card
- modelling code · transformers 5.18.0 modelling code, or the model repository's own modelling file at the pinned revision
Listed in the LLM Architecture Gallery checklist as “Solar Open 2 (250B)” (name only; see about).