Sarvam 105B
Sarvam AI · Sarvam · open weights
- MLA
- RoPE
- Partial RoPE
- Pre-norm
- QK-norm
- MoE
- Shared expert
- Dense first layers
Facts and where they come from
| Released | 2026-03 | config.jsonconfig.jsonHugging Face repository creation date (api.createdAt) |
|---|---|---|
| Licence | apache-2.0 | config.jsonconfig.jsonREADME metadata: license |
| Total parameters | 105B | labmodel cardmodel name sarvam-105b |
| Active parameters | 10.3B | labmodel cardREADME: 10.3B active parameters |
| Context length | 128K tokens | config.jsonconfig.jsonmax_position_embeddings |
| Norm placement | pre | codemodelling codetransformers 5.18.0 / repo modelling code for sarvam_mla: input_layernorm before attention, post_attention_layernorm before the MLP |
| Norm type | RMSNorm | codemodelling codetransformers 5.18.0 / repo modelling code for sarvam_mla: input_layernorm before attention, post_attention_layernorm before the MLP |
| QK-norm | yes | config.jsonconfig.jsonuse_qk_norm |
| Positional encoding | RoPE on 33.3% of each head | config.jsonconfig.jsonqk_rope_head_dim / (qk_nope_head_dim + qk_rope_head_dim): decoupled RoPE |
| Parallel attention and MLP | no | codemodelling codetransformers 5.18.0 / repo modelling code for sarvam_mla: input_layernorm before attention, post_attention_layernorm before the MLP |
Architecture, drawn from the data
MLA 64h, latent 512. 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) | 106B |
|---|---|
| Active per token (modelled) | 12.4B |
| Without embeddings and output head | 104B total, 10.3B active |
| Published weights (Hugging Face count) | 106B |
| KV cache per token, BF16 (layers that grow with context) | 36 KiB |
| KV cache + state at 128K tokens, BF16 | 4.5 GiB |
| Decode FLOPs per token at 4K context | 28.1 GFLOP |
| Prefill FLOPs for a 4K prompt | 95.1 TFLOP |
KV cache against context
Every architecture field
| Field | Value | Source |
|---|---|---|
| d_model | 4,096 | config.jsonconfig.jsonhidden_size |
| vocab | 262,144 | config.jsonconfig.jsonvocab_size |
| tied_embeddings | false | config.jsonconfig.jsontie_word_embeddings |
| mixers.mla.type | mla | codemodelling codemulti-head latent attention |
| mixers.mla.heads | 64 | config.jsonconfig.jsonnum_attention_heads |
| mixers.mla.q_lora_rank | null | codemodelling codeq_lora_rank: null (full-rank query) |
| mixers.mla.kv_lora_rank | 512 | config.jsonconfig.jsonkv_lora_rank |
| mixers.mla.qk_nope | 128 | config.jsonconfig.jsonqk_nope_head_dim |
| mixers.mla.qk_rope | 64 | config.jsonconfig.jsonqk_rope_head_dim |
| mixers.mla.v_head_dim | 128 | config.jsonconfig.jsonv_head_dim |
| ffns.dense.type | dense | codemodelling codeMLP block |
| ffns.dense.d_ff | 16,384 | config.jsonconfig.jsonintermediate_size |
| ffns.dense.gated | true | codemodelling codeMLP: gated (SwiGLU/GeGLU) |
| ffns.moe.type | moe | codemodelling codeMoE block |
| ffns.moe.experts | 128 | config.jsonconfig.jsonnum_experts |
| ffns.moe.active | 8 | config.jsonconfig.jsonnum_experts_per_tok |
| ffns.moe.d_expert | 2,048 | config.jsonconfig.jsonmoe_intermediate_size |
| ffns.moe.gated | true | codemodelling codeexperts are gated MLPs |
| ffns.moe.shared | 1 | config.jsonconfig.jsonnum_shared_experts |
| ffns.moe.d_shared | 2,048 | config.jsonconfig.jsonmoe_intermediate_size |
| layout | 1× mla/dense · 31× mla/moe | config.jsonconfig.jsonnum_hidden_layers, first_k_dense_replace |
Sources
- config.json @ abe264b
- model card
- announcement
- 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 “Sarvam (105B)” (name only; see about).