---
license: apache-2.0
base_model: sarvamai/sarvam-30b
tags:
  - gguf
  - sarvam
  - moe
  - indic
  - hindi
  - llama-cpp
language:
  - hi
  - en
  - ta
  - te
  - kn
  - ml
  - mr
  - bn
  - gu
  - pa
  - or
pipeline_tag: text-generation
quantized_by: mtrajan
---

# Sarvam-30B GGUF

GGUF quantizations of [sarvamai/sarvam-30b](https://huggingface.co/sarvamai/sarvam-30b) — the first publicly available GGUF for this model.

Created by applying [llama.cpp PR #20275](https://github.com/ggml-org/llama.cpp/pull/20275) which adds `sarvam_moe` architecture support to the converter and runtime.

> **Will this work with Ollama / LM Studio / Jan?**
>
> **Not yet.** These tools bundle mainline llama.cpp, which does not recognize `sarvam_moe`. You will see:
> ```
> error loading model: unknown model architecture: 'sarvam_moe'
> ```
> This GGUF **requires a patched llama.cpp** (with PR #20275 applied) until that PR merges into mainline. Once it does, Ollama / LM Studio / Jan will work automatically on their next update.
>
> To build a patched llama.cpp, use [mtr7x/sarvam-gguf](https://github.com/mtr7x/sarvam-gguf).

## Files

| File | Quant | Size | BPW | Notes |
|------|-------|------|-----|-------|
| sarvam-30b-q4_k_m.gguf | Q4_K_M | 19 GB | 4.87 | Recommended — good balance of quality and size |
| sarvam-30b-f16.gguf | F16 | 60 GB | 16.00 | Full precision, use for further quantization |

## How to use

### Option 1: Patch llama.cpp automatically (recommended)

```bash
git clone https://github.com/mtr7x/sarvam-gguf.git
cd sarvam-gguf
chmod +x patch_and_convert.sh
./patch_and_convert.sh
```

This clones llama.cpp, applies PR #20275, builds it, and you're ready to run.

### Option 2: Run with patched llama.cpp directly

```bash
./llama-cli \
    --model sarvam-30b-q4_k_m.gguf \
    --n-gpu-layers 99 \
    --ctx-size 2048 \
    --temp 0.7 \
    -no-cnv \
    --prompt "भारत के बारे में बताइए।"
```

### What does NOT work (yet)

| Tool | Status | Why |
|------|--------|-----|
| Ollama | `unknown model architecture` | Waiting on PR #20275 merge |
| LM Studio | `unknown model architecture` | Waiting on PR #20275 merge |
| Jan | `unknown model architecture` | Waiting on PR #20275 merge |
| llama.cpp (mainline) | `unknown model architecture` | PR #20275 not yet merged |
| llama.cpp (patched) | Works | This is what you need |

## Architecture

```
sarvamai/sarvam-30b
├── model_type: sarvam_moe
├── 30B params, 2.4B active
├── 19 layers (1 dense + 18 MoE)
├── 128 experts + 1 shared, top-6, sigmoid routing
├── 64 query heads, 4 KV heads, head_dim=64
├── vocab_size: 262,144 (Indic-optimized)
└── Apache 2.0
```

## Why this is needed

Sarvam open-sourced 30B and 105B under Apache 2.0, but mainline llama.cpp doesn't recognize `model_type: "sarvam_moe"` — the converter exits immediately. Contrary to what you might expect, **sigmoid routing is already supported** in llama.cpp (used by GLM4 and others). The actual blocker is a missing class registration + tensor mappings + C++ graph builder — all provided by PR #20275 (387 lines).

## The domino chain

```
PR #20275 merges into llama.cpp       ← pending
  → GGUF can be created               ← done (this repo)
    → Ollama updates its llama.cpp     ← blocked
      → Unsloth applies dynamic quants ← blocked
        → ollama run sarvam-30b        ← blocked
```

## Runtime support

| Runtime | Status |
|---------|--------|
| vLLM | PR #33942 merged |
| SGLang | Works |
| llama.cpp (patched) | Works (PR #20275) |
| llama.cpp (mainline) | Blocked — PR pending |
| Ollama | Blocked on llama.cpp |
| LM Studio | Blocked on llama.cpp |

## Credits

- [sarvamai](https://huggingface.co/sarvamai) for open-sourcing Sarvam-30B under Apache 2.0
- [sumitchatterjee13](https://github.com/sumitchatterjee13) for llama.cpp PR #20275
- Conversion pipeline: [mtr7x/sarvam-gguf](https://github.com/mtr7x/sarvam-gguf)

Read the full analysis: [Sarvam. Open is not sovereign](https://mtrajan.substack.com/p/sarvam-open-is-not-sovereign)