Providers

The model weights are the same across every row here — Qwen3.6-27B in some quantization. The variable is everything else: where the bytes get computed, which serving engine runs them, what sampling defaults the server applies, whether prefix-cache is hit, and how much round-trip latency the network adds.

Hostnames are abstracted to the substantive characteristics. The descriptive label captures engine + quantization + GPU/CPU + OS — enough to map to a known-good machine spec without leaking inventory.

Per-turn TTFT (first-token latency) and steady-state decode tok/s, bucketed by input-token length of that turn. We bucket per turn rather than per task because the agent loop's input grows monotonically inside a single task — buckets reveal how each provider scales prefill and decode under increasing context.

Buckets: 0–10k, 10–30k, 30–60k, 60–120k, 120k+ tokens. Buckets with fewer than 5 turns of data are dropped to avoid noise.

Read this as: a profile that holds steady across buckets has a working KV-cache / prefix-cache; a profile whose TTFT slopes up sharply is recomputing prefill on every turn.

Provider details below use public profile labels. They publish enough information to interpret the benchmark results: provider surface, runtime family, model or quantization, sampling policy, context limits, and broad hardware class for self-hosted profiles.

OpenRouter Qwen3.6-27B

• Profile: fiz-openrouter-qwen3-6-27b
• Surface: managed OpenAI-compatible API through OpenRouter.
• Model: qwen/qwen3-6-27b-instruct.
• Sampling: temperature=0.6, top_p=0.95, top_k=20, reasoning low.
• Limits: 128k advertised context, 32k max output.
• Cost profile: low cash cost per run; current medians put it near the budget end of the hosted profiles.
• Notes: provider-side routing and caching are opaque, so this profile is best read as the managed throughput reference for Qwen3.6-27B.

sindri-vllm

• Profile: sindri-vllm
• Surface: self-hosted vLLM on a local RTX-class CUDA workstation.
• Model: Qwen3.6-27B AutoRound int4.
• Sampling: temperature=0.6, top_p=0.95, top_k=20, reasoning low.
• Limits: 180k advertised context, 32k max output; effective usable context depends on runtime memory pressure.
• Notes: decode throughput is strong, but TTFT rises with long context. Prefix caching is the main performance lever for this profile.

sindri-llamacpp

• Profile: sindri-llamacpp
• Surface: self-hosted llama.cpp on the same local CUDA hardware class as sindri-vllm.
• Model: Qwen3.6-27B Q3_K_XL quantization.
• Sampling: temperature=0.6, top_p=0.95, top_k=20; provider-specific reasoning hints are not sent to llama.cpp.
• Notes: this profile isolates runtime and quantization differences against the same broad hardware class as the vLLM profile.

local-vllm-rtx3090

• Profile: local-vllm-rtx3090
• Surface: self-hosted vLLM on a mobile RTX-class CUDA system.
• Model: Qwen3.6-27B AutoRound int4.
• Sampling: same as sindri-vllm.
• Notes: this profile is wired up but not yet producing enough real reps for a comparable benchmark read.

local-omlx-qwen3-6-27b

• Profile: local-omlx-qwen3-6-27b
• Surface: self-hosted oMLX on an Apple silicon workstation class.
• Model: Qwen3.6-27B MLX 8-bit.
• Sampling: temperature=0.6, top_p=0.95, top_k=20, reasoning low.
• Limits: 128k advertised context, 32k max output.
• Notes: current results show slower TTFT and decode than the CUDA profiles at this model size.

local-rapidmlx-qwen3-6-27b

• Profile: local-rapidmlx-qwen3-6-27b
• Surface: self-hosted Rapid-MLX on an Apple silicon workstation class.
• Model: Qwen3.6-27B MLX 8-bit.
• Sampling: same as the oMLX profile.
• Notes: this profile is not yet producing comparable real reps.

local-lmstudio-qwen3-6-27b

• Profile: local-lmstudio-qwen3-6-27b
• Surface: LM Studio alternate runtime.
• Model: Qwen3.6-27B class local model.
• Notes: this profile is a placeholder until it produces real reps.