LLMBOTTLENECK.COM
sarvamai / sarvam_mla

sarvam-105b

106.0 billion parameters, routing 8 of 128 experts per token.

Open in the calculator →

Architecturepublished data
Quick answer
sarvam-105b needs about 64.9 GB at Q4_K_M — the format most people download — with 8,192 tokens of context. It fits entirely on 5 of 18 common devices, starting with the AMD Ryzen AI Max+ 395 with Radeon 8060S (128 GB). What else fits in 128 GB. No hardware for it? Rent a GPU that runs it, priced live.

Sized at 8,192 tokens of context with the whole model in device memory, across 18 common devices. Speeds are estimates, not benchmarks.

Find the AMD Ryzen AI Max+ 395 with Radeon 8060S, the smallest card listed here that holds it: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Latent (MLA) attention: head_dim 576 is kv_lora_rank 512 plus qk_rope_head_dim 64, the width of one compressed cache row, and q_head_dim 192 is qk_nope_head_dim 128 plus the same 64. The cache is sized from kv_lora_rank and qk_rope_head_dim like every other latent-attention model held here (audit 2026-09-18, A03). Licence declared on the repository: apache-2.0.

Weight-file size by format

FormatSizeRangeBasis
FP16210.5 GB207.9 GB – 213.1 GBreconstructed size
Q8_0111.9 GB110.5 GB – 113.3 GBreconstructed size
Q6_K86.4 GB85.3 GB – 87.5 GBreconstructed size
Q5_K_M74.7 GB73.8 GB – 75.7 GBreconstructed size
Q5_072.6 GB71.7 GB – 73.5 GBreconstructed size
Q4_K_M63.8 GB62.9 GB – 64.5 GBreconstructed size
Q4_059.5 GB58.7 GB – 60.3 GBreconstructed size
Q3_K_M52.2 GB48.2 GB – 56.3 GBreconstructed size
Q2_K38.6 GB36.7 GB – 40.4 GBreconstructed size

A file size is not the memory a run needs: the KV cache and the runtime reserve come on top, and the calculator adds both.

What the labels mean
published data
Read from a published source — a file's byte count, a model's configuration or a manufacturer's specification — or exact arithmetic on such values. Not a measurement on a machine.
reconstructed size
Weight size reconstructed from the pinned architecture, because no published file exists.
size range
Only a lower and an upper bound are claimed for this weight size.

Hardware ladder

Q4_K_M at 8,192 tokens · 5 of 18 devices hold it entirely
DeviceMemoryVerdictDecodeCalculator
NVIDIA GeForce RTX 507064.9 GB of 12 GB27 layers on system RAM~10 tok/s with offloadOpen →
AMD Radeon™ RX 9070 XT64.9 GB of 16 GB25 layers on system RAM~11 tok/s with offloadOpen →
NVIDIA GeForce RTX 408064.9 GB of 16 GB25 layers on system RAM~11 tok/s with offloadOpen →
NVIDIA GeForce RTX 508064.9 GB of 16 GB25 layers on system RAM~11 tok/s with offloadOpen →
NVIDIA GeForce RTX 5070 Ti64.9 GB of 16 GB25 layers on system RAM~11 tok/s with offloadOpen →
NVIDIA GeForce RTX 5060 Ti 16GB64.9 GB of 16 GB25 layers on system RAM~11 tok/s with offloadOpen →
NVIDIA GeForce RTX 409064.9 GB of 24 GB21 layers on system RAM~13 tok/s with offloadOpen →
AMD Radeon™ RX 7900 XTX64.9 GB of 24 GB21 layers on system RAM~13 tok/s with offloadOpen →
NVIDIA GeForce RTX 309064.9 GB of 24 GB21 layers on system RAM~13 tok/s with offloadOpen →
NVIDIA GeForce RTX 509064.9 GB of 32 GB17 layers on system RAM~16 tok/s with offloadOpen →
NVIDIA RTX 6000 Ada Generation64.9 GB of 48 GB9 layers on system RAM~26 tok/s with offloadOpen →
Apple M4 Pro64.9 GB of 64 GBdoes not fitnot calibratedOpen →
Apple M5 Pro64.9 GB of 64 GBdoes not fitnot calibratedOpen →
AMD Ryzen AI Max+ 395 with Radeon 8060S64.9 GB of 128 GBfits~29 tok/sAbout reading paceOpen →
Apple M3 Max64.9 GB of 128 GBfits~42 tok/sOpen →
Apple M4 Max64.9 GB of 128 GBfits~52 tok/sOpen →
NVIDIA DGX Spark64.9 GB of 128 GBfits~27 tok/sAbout reading paceOpen →
Apple M2 Ultra64.9 GB of 192 GBfits~67 tok/sOpen →

Decode is a calibrated estimate from the published calibration; capacity is the manufacturer's published ceiling, not guaranteed free memory. Fitting in memory is not the same as loading: whether the runtime and version you have supports this architecture and format on that machine has not been tested here. Offloaded rows assume there is enough system RAM for the overflow — the calculator checks that against the RAM you declare. Each “Open” link keeps this model, format and context.

Run it in the cloud

Rent a GPU that runs sarvam-105b

Every rentable machine that holds the whole model, from Vast.ai and RunPod, with this site's speed estimate for each and the provider's own price, read live. Cheapest first; sort by cost per token or speed instead, or switch the format.

Prices · 18:30 UTC, 22 Sept
18 configurations in stock · 14 out of stock
MachineSpeedPer hourPer M tokensWhereRent
8× RTX 306096 GBcheapest~217 tok/sFaster than you read$0.59$0.75Vast.aiMarketplace · 99.7% reliableRent on Vast.ai
4× RTX 309096 GB~326 tok/sFaster than you read$0.60$0.51Vast.aiMarketplace · 99.4% reliableRent on Vast.ai
4× RTX 5090128 GBrecommendedbest value~624 tok/sFaster than you read$1.07$0.48Vast.aiMarketplace · 99.6% reliableRent on Vast.ai
2× RTX 6000 Ada96 GB~181 tok/sFaster than you read$1.07$1.64Vast.aiMarketplace · 99.4% reliableRent on Vast.ai
1× RTX PRO 600096 GB~176 tok/sFaster than you read$1.14$1.79Vast.aiMarketplace · 98.8% reliableRunPod $1.69/h Rent on Vast.ai
1× A100 80GB PCIe80 GB~190 tok/sFaster than you read$1.19$1.73RunPodCommunity CloudRent on RunPod
2× RTX A600096 GB~145 tok/sFaster than you read$1.20$2.30Vast.aiMarketplace · 99.4% reliableRent on Vast.ai
4× L496 GB~104 tok/sFaster than you read$1.29$3.41Vast.aiMarketplace · 99.4% reliableRent on Vast.ai

First time renting a GPU? How it works, in four steps
  1. Create an account and add credit. Both providers are prepaid: $10 is enough to try any card on this page for hours. A new RunPod account that arrives through a referral link gets a one-time $5 credit when it first adds $10.
  2. Pick the card and the count shown here, and a template: llama.cpp or vLLM for an OpenAI-compatible API, or one with a web chat if you only want to talk to the model. “How to launch it” below gives the exact image and command.
  3. Wait for the download. The weights are fetched on the machine; tens of gigabytes take a few minutes on a data-centre connection. You pay by the second from the moment it starts.
  4. Stop it when you are done. A running machine bills even when idle. A stopped one costs nothing per hour, though its disk is usually still billed until you delete it.
How to launch it

No Q4_K_M file of sarvam-105b is catalogued here, so there is no exact command to vouch for. Start the machine from the llama.cpp server image ghcr.io/ggml-org/llama.cpp:server-cuda and point -hf at a Q4_K_M build from the publisher or a quantizer you trust — search Hugging Face for one. Check its file size against the memory figure on this page before you rent.

Every machine holds the whole model at Q4_K_M and 8,192 tokens of context, no offload. Speeds are this site’s single-stream decode estimates; prices are what each provider’s own API quoted, on-demand, for the whole machine. Vast hosts below 98% measured reliability are left out.

Referral links Vast.ai, RunPod and Novita pay us a share of what you spend if you sign up through these buttons. It costs you nothing, and it never decides an order or a recommendation: both are computed from the live price and the speed, and options that pay us nothing are listed and recommended on the same terms. How we rank

Buying a card instead, or paying by the token? Compare a year of sarvam-105b three ways →

Under the hood

Architecture, read from the publisher’s file

The numbers every figure above is computed from, with the file they came from.

Architecture

✓ Architecture read from the published config.json

Retrieved 2026-09-20 at pinned commit fb187764dbad; parameter count from the safetensors index at the same revision.

Architecture
sarvam_mla
Layers
32
Hidden size
4,096
Attention heads
64
Head dimension
576
Feed-forward width
16,384
Vocabulary
262,144
Context ceiling
131,072
RoPE theta
10,000
Experts
128
Experts / token
8
Expert width
2,048
Shared experts
1
Latent KV rank
512

Where the memory goes

tokenembedding× 32 decoder blocksLatent attentioncompressed KV rankfull contextRouted experts8 of 128 per tokenall residentoutputprojectiongrows with contextcapacity ≠ traffic

Latent attention projects keys and values into a compressed rank before caching them, which is why its KV cache is a fraction of a comparable dense model.

Open the exact official config.json ↗

More from sarvamai

sarvam-30b

Use this answer in your own product

As JSON, from the API
curl -s https://llmbottleneck.com/v1/analyze \
  -H "Authorization: Bearer $LLMB_KEY" -H "content-type: application/json" \
  -d '{"model":"sarvamai-sarvam-105b","quantization":"Q4_K_M","context":8192,"hardware":"nvidia-geforce-rtx-4090"}'

Same engine, same evidence, every field sourced. Free key in one step →

On your page, as a widget
<script src="https://llmbottleneck.com/widget.js"
  data-model="sarvamai-sarvam-105b" data-quantization="Q4_K_M"
  data-hardware="nvidia-geforce-rtx-4090" data-context="8192"></script>

No key needed. Unbranded, with your own buy button, on Pro and Business →

Citing this page

LLM Bottleneck. “sarvam-105b VRAM and hardware requirements.” Architecture from sarvamai/sarvam-105b at revision fb187764dbad, retrieved 2026-09-20. https://llmbottleneck.com/models/sarvamai-sarvam-105b

Every figure above is either the published value or a reconstruction whose measured error is on the accuracy page.

llmbottleneck
catalogue 2026-10-03models 327devices 135