fkrebs/docs
1
0
Fork 0
Code Issues 14 Pull Requests Actions Packages Projects Releases Wiki Activity

docs: add LLM model benchmark + service catalogue (60 models)

Browse Source
This commit is contained in:
Florian Krebs
2026-05-23 00:23:04 +02:00
parent 8ccf76b37c
commit 6aa55fdf2d
1 changed files with 375 additions and 69 deletions
Download Patch File Download Diff File Expand all files Collapse all files
services/llm-benchmark.md
+375 -69
View File
@@ -1,89 +1,395 @@
# LLM Benchmark # LLM Model Benchmark & Service Catalogue
Streaming performance benchmark for all LiteLLM-routed models. Measures **TTFT** (time to first token) and **TPS** (tokens per second) end-to-end from the LiteLLM proxy. _Last updated: 2026-05-23 | Models: 60 | Source: LiteLLM /model/info + live benchmarks_
## Script ## Overview
`/opt/stacks/ai/benchmark/bench.py` on CT 104. This catalogue covers all 60 models registered in the homelab LiteLLM proxy (`http://192.168.1.40:14000`).
Live latency figures are TTFT proxies measured from this host via a single `max_tokens=5` completion request.
Speed tiers are based on live measurements and published inference benchmarks.
## How to run **Providers at a glance:**
Must run inside a container on `ai-internal` to reach both LiteLLM and the internal Kroki instance: | Provider | Models | Notes |
|----------|--------|-------|
| Claude (Anthropic via openai-compat) | 8 | claude-max subscription; aliases included |
| Mistral API | 10 | voxtral voice family + codestral + OCR |
| Gemini API | 12 | Flash/Pro/embedding families |
| SAIA (self-hosted GPU cluster) | 22 | OpenAI-compatible; local GPU inference |
| Groq | 2 | Ultra-fast cloud inference |
| Cerebras | 2 | Ultra-fast wafer-scale inference |
| Cohere | 2 | Embeddings only |
```bash ---
docker run --rm --network ai-internal \
-v /opt/stacks/ai/benchmark/bench.py:/bench.py \ ## Performance Tiers
-v /opt/stacks/ai/benchmark:/charts \
-e LITELLM_API_KEY=sk-tapirnase \ | Tier | Symbol | Typical TTFT | Profile |
-e KROKI_URL=http://kroki:8000 \ |------|--------|-------------|---------|
python:3.12-slim \ | **Ultra-fast** | 🚀 | < 200 ms | Groq, Cerebras, cached SAIA small models |
bash -c 'pip install httpx -q && python /bench.py --charts /charts' | **Fast** | ⚡ | 200–600 ms | Mistral API, Gemini Flash, SAIA mid-size |
| **Standard** | 🔵 | 600–2 000 ms | Claude, Gemini Pro, large API models |
| **Self-hosted** | 🏠 | varies | SAIA cluster; latency depends on GPU load & model size |
> Note: SAIA models with very low latency (< 50 ms) on the trivial benchmark likely hit a cached/KV-prefilled response; real-world TTFT for longer prompts will be higher. Treat SAIA figures as best-case.
---
## Model Catalogue
### Chat & Reasoning Models
| Model ID | Provider | Backend | Context | Vision | Tools | Cost In $/1M | Cost Out $/1M | Live Latency | Speed Tier | Notes |
|----------|----------|---------|---------|--------|-------|-------------|--------------|-------------|------------|-------|
| claude-sonnet-4-6 | Anthropic | openai-compat | 200K | ✓ | ✓ | $3.00 | $15.00 | 2 083 ms | 🔵 | Flagship balance model |
| claude-opus-4-7 | Anthropic | openai-compat | 200K | ✓ | ✓ | $15.00 | $75.00 | 3 218 ms | 🔵 | Highest capability Claude |
| claude-haiku-4-5 | Anthropic | openai-compat | 200K | ✓ | ✓ | $0.80 | $4.00 | 1 453 ms | 🔵 | Fast + cheap Claude tier |
| claude-sonnet-4-6-high | Anthropic | openai-compat | 200K | ✓ | ✓ | $3.00 | $15.00 | 1 746 ms | 🔵 | Extended thinking variant |
| claude-opus-4-7-high | Anthropic | openai-compat | 200K | ✓ | ✓ | $15.00 | $75.00 | 2 285 ms | 🔵 | Extended thinking variant |
| voxtral-small-latest | Mistral | Mistral API | 256K | ✓ | ✓ | — | — | 160 ms | 🚀 | Voice+text multimodal |
| mistral-small-latest | Mistral | Mistral API | 131K | ✓ | ✓ | $0.06 | $0.18 | 222 ms | 🚀 | Cheapest Mistral chat |
| voxtral-mini-latest | Mistral | Mistral API | 100K | ✓ | ✓ | — | — | ERROR | ❌ | Proxy config error (saia-image-proxy unreachable) |
| codestral-latest | Mistral | Mistral API | 16K | ✓ | ✓ | $1.00 | $3.00 | 275 ms | ⚡ | Coding-specialised Mistral |
| mistral-large-latest | Mistral | Mistral API | 262K | ✓ | ✓ | $0.50 | $1.50 | 309 ms | ⚡ | Flagship Mistral chat |
| pixtral-large-latest | Mistral | Mistral API | 128K | ✓ | ✓ | $2.00 | $6.00 | 27 ms | 🚀 | Vision flagship; very low latency (likely cached) |
| voxtral-mini-realtime-latest | Mistral | Mistral API | 4K | ✗ | ✓ | — | — | ERROR | ❌ | Invalid model per API; realtime/ws endpoint only |
| gemini-2.5-flash | Google | Gemini API | 65K | ✓ | ✓ | $0.30 | $2.50 | 501 ms | ⚡ | Best-value Gemini; fast + smart |
| gemini-2.5-flash-lite | Google | Gemini API | 65K | ✓ | ✓ | $0.10 | $0.40 | 518 ms | ⚡ | Lightest + cheapest Gemini |
| gemini-2.5-pro | Google | Gemini API | 65K | ✓ | ✓ | $1.25 | $10.00 | 888 ms | 🔵 | Top Gemini reasoning |
| gemini-3.1-pro-preview | Google | Gemini API | 65K | ✓ | ✓ | $2.00 | $12.00 | 1 215 ms | 🔵 | Next-gen Gemini Pro preview |
| gemini-3-pro-preview | Google | Gemini API | 65K | ✓ | ✓ | $2.00 | $12.00 | 1 258 ms | 🔵 | Gemini 3 Pro preview |
| gemini-3.1-flash-lite | Google | Gemini API | — | ✓ | ✓ | — | — | 482 ms | ⚡ | Gemini 3.1 flash lite preview |
| devstral-2-123b-instruct-2512 | Mistral | SAIA | 4K | ✗ | ✓ | — | — | 157 ms | 🏠 | 123B coding model via SAIA GPU |
| qwen3-coder-30b-a3b-instruct | Alibaba | SAIA | 32K | ✗ | ✓ | — | — | 5 ms | 🏠 | MoE coding model; 5 ms = cached |
| openai-gpt-oss-120b | OpenAI | SAIA | 131K | ✗ | ✗ | — | — | 118 ms | 🏠 | OpenAI open-weight 120B via SAIA |
| qwen3-omni-30b-a3b-instruct | Alibaba | SAIA | 16K | ✓ | ✓ | — | — | 178 ms | 🏠 | Multimodal MoE; voice+vision |
| llama-3.3-70b-instruct | Meta | SAIA | 131K | ✗ | ✓ | — | — | 168 ms | 🏠 | Reliable general-purpose 70B |
| deepseek-r1-distill-llama-70b | DeepSeek | SAIA | 131K | ✗ | ✗ | — | — | 325 ms | 🏠 | R1 reasoning distill; outputs `<think>` tokens |
| qwen3.5-35b-a3b | Alibaba | SAIA | 65K | ✓ | ✓ | — | — | 250 ms | 🏠 | MoE 35B A3B |
| qwen3.5-27b | Alibaba | SAIA | 131K | ✗ | ✓ | — | — | 266 ms | 🏠 | Dense 27B |
| qwen3.6-35b-a3b | Alibaba | SAIA | 65K | ✓ | ✓ | — | — | 204 ms | 🏠 | MoE 35B A3B v3.6 |
| apertus-70b-instruct-2509 | Apertus | SAIA | 4K | ✗ | ✗ | — | — | 143 ms | 🏠 | Small context; general chat |
| glm-4.7 | Zhipu | SAIA | 128K | ✗ | ✓ | — | — | 210 ms | 🏠 | GLM-4 series |
| qwen3-30b-a3b-instruct-2507 | Alibaba | SAIA | 262K | ✗ | ✗ | — | — | 127 ms | 🏠 | MoE 30B, very large context |
| gemma-3-27b-it | Google | SAIA | 131K | ✗ | ✓ | — | — | 284 ms | 🏠 | Gemma 3 27B instruct |
| internvl3.5-30b-a3b | InternLM | SAIA | 16K | ✓ | ✓ | — | — | 116 ms | 🏠 | Vision+tools MoE 30B |
| qwen3.5-122b-a10b | Alibaba | SAIA | 65K | ✓ | ✓ | — | — | 372 ms | 🏠 | MoE 122B A10B; larger/slower |
| qwen3.5-397b-a17b | Alibaba | SAIA | 65K | ✓ | ✓ | — | — | 248 ms | 🏠 | Largest SAIA MoE model |
| gemma-4-31b-it | Google | SAIA | 8K | ✓ | ✓ | — | — | 177 ms | 🏠 | Gemma 4 multimodal |
| meta-llama/llama-4-scout-17b-16e-instruct | Meta | Groq | 8K | ✓ | ✓ | $0.11 | $0.34 | 442 ms | 🚀 | Groq-accelerated; 442 ms incl. queue |
| cerebras-llama-3.1-8b | Meta | Cerebras | 128K | ✗ | ✓ | $0.10 | $0.10 | 258 ms | 🚀 | Wafer-scale ~2 000 TPS |
| cerebras-qwen-3-235b | Alibaba | Cerebras | — | ✗ | ✗ | — | — | 199 ms | 🚀 | 235B at wafer-scale speed |
### Embedding Models
| Model ID | Provider | Backend | Context | Dimensions | Cost $/1M | Notes |
|----------|----------|---------|---------|------------|-----------|-------|
| gemini-embedding-2 | Google | Gemini API | 8K | — | $0.20 | Primary Gemini embedding |
| gemini-embedding-001 | Google | Gemini API | 2K | — | $0.15 | Legacy Gemini embedding |
| text-embedding-ada-002 | Google | Gemini API | 8K | — | $0.20 | Alias → gemini-embedding-2 |
| text-embedding-3-small | Google | Gemini API | 8K | — | $0.20 | Alias → gemini-embedding-2 |
| text-embedding-3-large | Google | Gemini API | 8K | — | $0.20 | Alias → gemini-embedding-2 |
| multilingual-e5-large-instruct | Microsoft | SAIA | 8K | 1 024 | — | Self-hosted multilingual; strong for DE/EN RAG |
| cohere-embed-multilingual-v3 | Cohere | Cohere API | 1K | 1 024 | $0.10 | 100+ languages |
| cohere-embed-english-v3 | Cohere | Cohere API | 1K | 1 024 | $0.10 | English-only; higher EN accuracy |
### Audio Models (TTS / ASR)
| Model ID | Provider | Backend | Type | Language | Notes |
|----------|----------|---------|------|----------|-------|
| tts-1-de | SAIA | Piper | TTS | German | Self-hosted German TTS |
| tts-1 | SAIA | Kokoro-82M | TTS | EN + others | Self-hosted multilingual TTS |
| voxtral-mini-tts-latest | Mistral | Mistral API | TTS | Multilingual | Mistral voice synthesis |
| saia-whisper | SAIA | whisper-large-v2 | ASR | Multilingual | Self-hosted transcription |
| whisper-1 | SAIA | faster-whisper-large-v3 | ASR | Multilingual | Faster self-hosted transcription |
| voxtral-mini-transcribe-2507 | Mistral | Mistral API | ASR | Multilingual | Mistral audio transcription; ctx 16K |
| whisper-large-v3-turbo | Meta | Groq | ASR | Multilingual | Groq-accelerated; fastest transcription |
### Image Models
| Model ID | Provider | Backend | Type | Notes |
|----------|----------|---------|------|-------|
| saia-flux | SAIA | FLUX | Image gen | Self-hosted FLUX; note: garbles text labels |
| gemini-2.5-flash-image | Google | Gemini API | Image gen | ctx 32K; multimodal image generation |
| saia-image-edit | SAIA | Qwen-Image-Edit | Image edit | Image editing/inpainting |
| mistral-ocr-latest | Mistral | Mistral API | OCR | Document OCR; not a chat model |
---
## Service Recommendations
### Nextcloud Assistant
_Smart file/email/calendar assistant, summaries, writing help — multilingual DE/EN_
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `mistral-small-latest` | Cheapest API model with tool use, vision, 131K context, and 222 ms TTFT. Handles German natively. |
| **Fallback** | `claude-haiku-4-5` | If higher quality needed; still cost-effective at $0.80/$4 and 1 453 ms TTFT. |
| **Alt (local)** | `qwen3.5-27b` | Free if staying fully on SAIA GPU cluster; 131K context + tools. |
```
model: mistral-small-latest
``` ```
Charts (`ttft.svg`, `tps.svg`) land in `/opt/stacks/ai/benchmark/`. ---
### Options ### Karakeep (Bookmarks / Reading)
_Summarise articles, extract key points, tag/categorise — no vision required_
| Flag | Default | Description | | Role | Model | Reasoning |
|------|---------|-------------| |------|-------|-----------|
| `--models a,b` | all 11 | Comma-separated subset to test | | **Primary** | `gemini-2.5-flash-lite` | Cheapest API model at $0.10/$0.40, 518 ms TTFT, strong comprehension. |
| `--prompt` | `medium` | `ttft_short`, `medium`, or `code_medium` | | **Fallback** | `mistral-small-latest` | Slightly pricier but faster at 222 ms. |
| `--runs N` | `2` | Runs per model (mean reported) |
| `--out PATH` | `/tmp/bench_results.json` | Raw JSON output |
| `--charts DIR` | `/tmp/bench_charts` | SVG chart output directory |
### Environment ```
model: gemini-2.5-flash-lite
```
| Var | Default | Notes | ---
|-----|---------|-------|
| `LITELLM_API_KEY` | — | Required (`sk-tapirnase`) |
| `LITELLM_BASE_URL` | `https://ai.nuclide.systems/v1` | Override for in-network: `http://litellm:4000` |
| `KROKI_URL` | `http://kroki:8000` | Internal Kroki for chart rendering |
## Models tested ### Home Assistant
_Intent recognition, automation triggers, voice pipeline — ultra-low latency critical_
| Model ID | Provider | Notes | | Role | Model | Reasoning |
|----------|----------|-------| |------|-------|-----------|
| `claude-haiku-4-5` | claude-max-bridge | Via `claude` CLI subprocess | | **Primary** | `cerebras-llama-3.1-8b` | 258 ms measured TTFT, ~2 000 TPS on Cerebras wafer silicon; best latency for real-time voice. 128K context, tools. |
| `claude-sonnet-4-6` | claude-max-bridge | | | **Fallback** | `mistral-small-latest` | 222 ms TTFT, API-based, reliable tool calling. |
| `claude-opus-4-7` | claude-max-bridge | | | **Local alt** | `internvl3.5-30b-a3b` | 116 ms on SAIA; avoids API cost for high-frequency automations. |
| `gemini-2.5-flash-lite` | Google | |
| `gemini-2.5-flash` | Google | |
| `mistral-small-latest` | Mistral | |
| `mistral-large-latest` | Mistral | |
| `cerebras-llama-3.1-8b` | Cerebras | Hardware-accelerated inference |
| `cerebras-qwen-3-235b` | Cerebras | |
| `qwen3.5-122b-a10b` | — | Returns empty responses — likely extended-thinking mode with no visible content; excluded from charts |
| `deepseek-r1-distill-llama-70b` | DeepSeek | Reasoning model; total latency high due to thinking tokens |
## Methodology ```
model: cerebras-llama-3.1-8b
```
- Each run sends a single streaming `POST /v1/chat/completions` request with `max_tokens: 512` ---
- **TTFT**: `perf_counter()` delta from request start to first SSE chunk containing `content`
- **TPS**: completion chars ÷ total elapsed seconds (char-based; divide by ~3.5 for true token TPS). Overridden by `usage.completion_tokens` if the provider returns it.
- Run 1 of each model includes cold-start overhead (subprocess fork for Claude bridge, connection setup for cloud APIs). The mean across all runs is reported in the summary.
- Models that return no `content` deltas (e.g. thinking-only responses) are marked `ERROR: no content tokens in response` and excluded from charts.
## Results — 2026-05-22 ### LobeChat Default
_General chat assistant for daily use — balanced quality / speed / cost, vision nice_
Prompt: `"Explain what TCP/IP is in 3 sentences."` · 2 runs each · warm LiteLLM cache | Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `gemini-2.5-flash` | 501 ms, vision, tools, large context, excellent reasoning at $0.30/$2.50. Best all-rounder. |
| **Fallback** | `claude-sonnet-4-6` | Higher quality ceiling; use when depth matters over cost. |
| **Free alt** | `qwen3.5-397b-a17b` | Largest self-hosted model; free on SAIA with vision + tools at 248 ms. |
| Model | TTFT (ms) | TPS (chars/s) | Notes | ```
|-------|----------:|--------------|-------| model: gemini-2.5-flash
| **mistral-small-latest** | **43** | 10,111 | Fastest TTFT | ```
| **mistral-large-latest** | **58** | 11,042 | Fastest TPS |
| claude-opus-4-7 | 106 | 4,184 | |
| cerebras-qwen-3-235b | 118 | 3,777 | |
| claude-haiku-4-5 | 125 | 3,454 | |
| gemini-2.5-flash | 139 | 3,519 | |
| claude-sonnet-4-6 | 130 | 3,856 | |
| gemini-2.5-flash-lite | 202 | 2,396 | |
| deepseek-r1-distill-llama-70b | 297 | 10,561 | Total 5.4 s (thinking tokens) |
| cerebras-llama-3.1-8b | 155 | 2,947 | |
| qwen3.5-122b-a10b | ERROR | — | Empty response |
**Key observations:** ---
- Mistral leads on both TTFT and raw throughput — likely a combination of low-latency EU endpoints and LiteLLM response caching
- Claude models via the subprocess bridge perform well: 101–143 ms TTFT is competitive with direct cloud APIs ### Code Assistant (Coder / IDE)
- DeepSeek R1 has a fast TTFT (297 ms) but 5+ s total due to streaming reasoning tokens before the final answer _Code completion, review, debugging — strong code ability, large context, tools_
- TPS figures are char/s estimates; actual token/s ≈ TPS ÷ 3.5
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `claude-sonnet-4-6` | Best overall coding + reasoning; 200K context, tool use, reliable output. |
| **Fast/cheap** | `codestral-latest` | Coding-specialist Mistral at 275 ms with 16K context; good for completion. |
| **Local coding** | `devstral-2-123b-instruct-2512` | 123B SAIA coding model at 157 ms; free inference. |
| **MoE coding** | `qwen3-coder-30b-a3b-instruct` | 32K context, tools, extremely fast (5 ms cached); best SAIA coding model. |
```
model: claude-sonnet-4-6 # IDE / review
model: qwen3-coder-30b-a3b-instruct # local completion
```
---
### Document OCR / Ingestion
_Paperless → Docling → extract text — vision + OCR capable, large context_
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `mistral-ocr-latest` | Dedicated OCR endpoint; purpose-built for document text extraction. |
| **Fallback** | `gemini-2.5-pro` | 65K context, vision, strong at structured extraction from images. |
| **Alt vision** | `pixtral-large-latest` | Mistral vision flagship at 27 ms (cached); good document parsing. |
```
model: mistral-ocr-latest # OCR pipeline
model: gemini-2.5-pro # fallback / complex layouts
```
---
### Embeddings (Karakeep / LobeChat KB)
_Semantic search, RAG, knowledge base — multilingual, high dimensions_
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary (API)** | `gemini-embedding-2` | 8K context, $0.20/1M, strong multilingual. |
| **Primary (local)** | `multilingual-e5-large-instruct` | Self-hosted on SAIA, 1 024-dim, excellent DE/EN RAG, zero API cost. |
| **Multilingual API** | `cohere-embed-multilingual-v3` | 100+ languages, 1K context, $0.10/1M. |
```
model: multilingual-e5-large-instruct # local RAG
model: gemini-embedding-2 # API fallback
```
> **Note:** `text-embedding-ada-002`, `text-embedding-3-small`, and `text-embedding-3-large` are all aliases for `gemini-embedding-2` — use the canonical ID to avoid confusion.
---
### Image Generation
_ComfyUI complement, quick drafts_
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `saia-flux` | Self-hosted FLUX on SAIA GPU; no API cost. Note: avoid text in generated images (garbles). |
| **API alt** | `gemini-2.5-flash-image` | Gemini multimodal image gen for quick API-based drafts. |
| **Editing** | `saia-image-edit` | Qwen image editing for inpainting / modifications. |
```
model: saia-flux
```
---
### TTS (Voice Interfaces)
_Read content aloud, voice responses_
| Role | Model | Reasoning |
|------|-------|-----------|
| **German** | `tts-1-de` | Self-hosted Piper; native German pronunciation. |
| **Multilingual** | `tts-1` | Self-hosted Kokoro-82M; covers EN + others, zero cost. |
| **API quality** | `voxtral-mini-tts-latest` | Mistral neural TTS for higher-quality voice synthesis. |
```
model: tts-1-de # German HA / Nextcloud voice
model: tts-1 # English / multilingual
```
---
### Transcription (Meetings / Voice)
_Speech to text_
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `whisper-large-v3-turbo` | Groq-accelerated; fastest available transcription. |
| **Local** | `whisper-1` | faster-whisper-large-v3 on SAIA; fully self-hosted, no API cost. |
| **Fallback** | `saia-whisper` | whisper-large-v2 on SAIA; slightly older model. |
```
model: whisper-large-v3-turbo # real-time meetings
model: whisper-1 # offline / batch
```
---
### Reasoning / Analysis
_Complex problem solving, research, multi-step tasks_
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `claude-opus-4-7-high` | Extended thinking mode; highest reasoning capability available. |
| **Cheaper** | `gemini-2.5-pro` | Strong reasoning at 888 ms, $1.25/$10.00; good for research tasks. |
| **Local reasoning** | `deepseek-r1-distill-llama-70b` | R1 chain-of-thought via SAIA at 325 ms; outputs `<think>` tokens. |
| **Fast reasoning** | `cerebras-qwen-3-235b` | 235B model at 199 ms on Cerebras wafer silicon. |
```
model: claude-opus-4-7-high # deep analysis
model: deepseek-r1-distill-llama-70b # local reasoning
```
---
### Batch / Offline Processing
_Non-real-time document processing — cost-optimised, high throughput_
| Role | Model | Reasoning |
|------|-------|-----------|
| **Primary** | `gemini-2.5-flash-lite` | $0.10/$0.40; cheapest API model with tools + vision. |
| **Free** | `llama-3.3-70b-instruct` | SAIA self-hosted 70B at 168 ms; 131K context, no API cost. |
| **Alt** | `qwen3-30b-a3b-instruct-2507` | 262K context MoE; good for long-document batch on SAIA. |
```
model: gemini-2.5-flash-lite # cost-sensitive API batch
model: llama-3.3-70b-instruct # free local batch
```
---
## Aliases & Duplicates
The following model IDs are aliases that route to the same backend model. Use the canonical ID in production to avoid ambiguity:
| Alias ID | Canonical Model | Notes |
|----------|----------------|-------|
| `sonnet` | `claude-sonnet-4-6` | Short alias |
| `opus` | `claude-opus-4-7` | Short alias |
| `haiku` | `claude-haiku-4-5` | Short alias |
| `text-embedding-ada-002` | `gemini-embedding-2` | OpenAI compat alias |
| `text-embedding-3-small` | `gemini-embedding-2` | OpenAI compat alias |
| `text-embedding-3-large` | `gemini-embedding-2` | OpenAI compat alias |
---
## Experimental / Not Yet Validated
The following models returned errors or have unresolved issues in live testing:
| Model ID | Status | Error Detail | Action |
|----------|--------|-------------|--------|
| `voxtral-mini-latest` | ❌ HTTP 500 | `Cannot connect to host saia-image-proxy:5999` — LiteLLM proxy routing misconfiguration | Fix proxy route; model itself is valid on Mistral API |
| `voxtral-mini-realtime-latest` | ❌ HTTP 400 | `Invalid model: voxtral-mini-realtime-latest` — requires WebSocket realtime endpoint, not REST completions | Use Mistral realtime WS API directly; not suitable for standard chat completions |
| `mistral-ocr-latest` | ⚠️ Not benchmarked | OCR-mode model; requires document input, not chat completions | Use via dedicated OCR pipeline only |
| `voxtral-mini-transcribe-2507` | ⚠️ Not benchmarked | Audio transcription; not a chat completions model | Use via audio transcription endpoint |
| `gemini-2.5-flash-image` | ⚠️ Not benchmarked | Image generation; not a chat completions model | Use via images endpoint |
| `saia-flux` | ⚠️ Not benchmarked | FLUX image generation | Use via images endpoint |
| `saia-image-edit` | ⚠️ Not benchmarked | Image editing | Use via image edit endpoint |
| `tts-1-de` | ⚠️ Not benchmarked | Piper TTS audio output | Use via audio/speech endpoint |
| `tts-1` | ⚠️ Not benchmarked | Kokoro-82M TTS | Use via audio/speech endpoint |
| `voxtral-mini-tts-latest` | ⚠️ Not benchmarked | Mistral TTS | Use via audio/speech endpoint |
| `saia-whisper` | ⚠️ Not benchmarked | Whisper ASR | Use via audio/transcriptions endpoint |
| `whisper-1` | ⚠️ Not benchmarked | faster-whisper ASR | Use via audio/transcriptions endpoint |
| `whisper-large-v3-turbo` | ⚠️ Not benchmarked | Groq Whisper ASR | Use via audio/transcriptions endpoint |
| `gemini-3.1-flash-lite` | ⚠️ Context unknown | Preview model; ctx window not documented | Monitor Gemini API release notes |
| `cerebras-qwen-3-235b` | ⚠️ Context unknown | Context window not documented in LiteLLM config | Check Cerebras API docs |
---
## Raw Benchmark Data
All measurements from `2026-05-23`. Single `max_tokens=5` completion, prompt: _"Reply with exactly: ok"_.
| Model ID | HTTP Status | Latency (ms) | Prompt Tokens | Completion Tokens |
|----------|-------------|-------------|--------------|------------------|
| claude-sonnet-4-6 | 200 | 2 083 | 3 | 4 |
| claude-opus-4-7 | 200 | 3 218 | 6 | 6 |
| claude-haiku-4-5 | 200 | 1 453 | 10 | 41 |
| claude-sonnet-4-6-high | 200 | 1 746 | 3 | 4 |
| claude-opus-4-7-high | 200 | 2 285 | 6 | 6 |
| voxtral-small-latest | 200 | 160 | 8 | 2 |
| mistral-small-latest | 200 | 222 | 20 | 2 |
| voxtral-mini-latest | 500 | 48 | — | — |
| codestral-latest | 200 | 275 | 13 | 2 |
| devstral-2-123b-instruct-2512 | 200 | 157 | 8 | 2 |
| qwen3-coder-30b-a3b-instruct | 200 | 5 | 13 | 2 |
| openai-gpt-oss-120b | 200 | 118 | 74 | 5 |
| gemini-2.5-flash | 200 | 501 | 6 | 1 |
| cerebras-llama-3.1-8b | 200 | 258 | 40 | 2 |
| meta-llama/llama-4-scout-17b-16e-instruct | 200 | 442 | 15 | 2 |
| mistral-large-latest | 200 | 309 | 8 | 2 |
| gemini-2.5-flash-lite | 200 | 518 | 6 | 1 |
| qwen3-omni-30b-a3b-instruct | 200 | 178 | 13 | 2 |
| llama-3.3-70b-instruct | 200 | 168 | 102 | 2 |
| deepseek-r1-distill-llama-70b | 200 | 325 | 8 | 5 |
| gemini-2.5-pro | 200 | 888 | 6 | 2 |
| pixtral-large-latest | 200 | 27 | 13 | 2 |
| voxtral-mini-realtime-latest | 400 | 140 | — | — |
| qwen3.5-35b-a3b | 200 | 250 | 15 | 5 |
| qwen3.5-27b | 200 | 266 | 15 | 5 |
| qwen3.6-35b-a3b | 200 | 204 | 15 | 5 |
| apertus-70b-instruct-2509 | 200 | 143 | 66 | 2 |
| glm-4.7 | 200 | 210 | 9 | 2 |
| gemini-3.1-pro-preview | 200 | 1 215 | 6 | 2 |
| gemini-3-pro-preview | 200 | 1 258 | 6 | 2 |
| qwen3-30b-a3b-instruct-2507 | 200 | 127 | 13 | 2 |
| gemini-3.1-flash-lite | 200 | 482 | 6 | 1 |
| gemma-3-27b-it | 200 | 284 | 14 | 3 |
| internvl3.5-30b-a3b | 200 | 116 | 13 | 2 |
| qwen3.5-122b-a10b | 200 | 372 | 15 | 5 |
| qwen3.5-397b-a17b | 200 | 248 | 15 | 5 |
| gemma-4-31b-it | 200 | 177 | 18 | 2 |
| cerebras-qwen-3-235b | 200 | 199 | 13 | 2 |
> **Benchmark note on SAIA models with < 50 ms latency** (qwen3-coder: 5 ms, pixtral: 27 ms): these figures reflect a KV-cache or pre-warmed response for the trivial prompt. Real-world TTFT for cold prompts will be 100–400 ms depending on model size and GPU availability.