LiteLLM vs vLLM

Auto Router: 45% Lower Cost on 25 SWE-bench Tasks

We solved 23 of 25 SWE-bench Verified tasks with LiteLLM's experimental capability router for $11.15, compared with $20.27 using Opus 5.

Introducing LiteLLM Fusion: 56% More Tasks Solved Than Fable 5

LiteLLM Auto Router Fusion ran three models on the same task and synthesized their work, solving 14 of 21 Terminal-Bench tasks against 9 for Claude Fable-5 alone. Total spend rose 36%, cost per solved task fell 12%, and turn latency went up 5x.

Configure LiteLLM as Codex Model Provider

A developer guide demonstrating how to route Codex to use LiteLLM as a custom model provider. The post explains exposing a LiteLLM API key as an OS environment variable, updating Codex's .codex/config.toml to set model_provider to 'litellm' and add provider-specific fields (base_url, env_key, wire_api, streaming options), setting optional custom HTTP headers, and noting that session models are fixed at session creation. The author also advises verifying usage via LiteLLM dashboard logs and links to LiteLLM and Codex documentation for reference.

• Author Julia Shevchenko published a how-to on dev.to on 2026-08-28 about configuring LiteLLM as a gateway for Codex.
• LiteLLM exposes an OpenAI-compatible interface and can act as a gateway for LLMs.
• Required steps include setting LITELLM_API_KEY as an OS environment variable and updating .codex/config.toml to set model_provider = "litellm" and provider-specific settings (base_url, env_key, wire_api, streaming options).

MiniMax H3 on vLLM-Omni: From System-Wide Optimization to Real-Time Serving with FastVideo’s FastH3

How vLLM-Omni optimizes and scales the complete MiniMax H3 stack, then integrates FastVideo’s four-step FastH3 for generation faster than playback.

Qwen3-8B inference benchmark and FP8 on Blackwell

Independent benchmarks compare Qwen3-8B inference on an RTX PRO 6000 Blackwell (96 GB) across three serving stacks (vLLM 0.27.1, SGLang 0.5.9, and llama.cpp CUDA). At concurrency 32 using BF16, vLLM achieved 1,725 aggregate tokens/s (TTFT p50 39 ms), SGLang 1,327 tok/s (TTFT p50 42 ms), and llama.cpp 428 tok/s (TTFT p50 316 ms). Applying an FP8 checkpoint to vLLM increased throughput by ~1.5x (aggregate 1,725 -> 2,597 tok/s; single-stream 86 -> 130 tok/s) with lower latency and no detected regressions on a fixed factual check. The author documents methodology, reproductions, and an sm_120-specific kernel workaround required to run FP8 on workstation Blackwell hardware.

• GPU used: RTX PRO 6000 Blackwell, 96 GB (workstation Blackwell, sm_120).
• Model benchmarked: Qwen3-8B across vLLM 0.27.1, SGLang 0.5.9, and llama.cpp (CUDA).
• BF16, concurrency 32 aggregate throughput: vLLM 1,725 tok/s; SGLang 1,327 tok/s; llama.cpp 428 tok/s.

Tokens-per-Second Benchmarks Explained

This technical guide explains what "tokens per second" (tok/s) actually measures for local LLM inference, why single-user tok/s numbers can be misleading, and how concurrency, batching, and prompt processing change the observed speed. It contrasts single-user latency with server throughput, highlights vLLM's continuous-batching advantage versus Ollama under high concurrency, defines related metrics (P99 latency, time to first token / TTFT), and provides practical measurement advice using tools like Ollama and vLLM and calculators from notAcalculator. The article also gives realistic tok/s expectations for different model sizes on consumer hardware and lists practical tips for reading and running benchmarks yourself.

• Tokens are the unit of both billing and speed for LLMs; tokenization affects cost and measured tok/s.
• Under a Red Hat benchmark on an A100 40GB with Llama 3.1 8B, vLLM peaked around 793 tok/s combined throughput versus about 41 tok/s for Ollama at high concurrency (~19x gap).
• vLLM's key innovation is continuous batching (plus PagedAttention), which increases total throughput under concurrency compared with single-request processing tools.