230 Models, One API: The Nous Portal TTFT Benchmark

Author: Agentic Architect Date: 2026–07–27

The Problem Nobody Else Solved

If you’ve ever tried to pick the right LLM for a task, you know the drill: check OpenRouter rankings, cross-reference Artificial Analysis, scan Chatbot Arena, then guess.

Here’s what nobody gives you: • TTFT on a single API endpoint — not cross-provider averages, not synthetic benchmarks • Agent readiness — which models actually work in tool loops vs pure chat • Specialty cross-analysis — how coding models compare to reasoning models on speed • Free pool performance — not just “it’s free” but “how fast is it actually”

So I built it.

Methodology

All 289 models on Nous Portal (https://portal.nousresearch.com) were tested with the same Chinese prompt on the same machine, same API endpoint. Each model got 3 streaming SSE runs; TTFT (time-to-first-token) was averaged.

Prompt: “列出台灣 2025 年 AI 發展的 3 個重大事件。”

Why Chinese? Because most benchmarks use English, and TTFT is sensitive to tokenization — Chinese text produces different token counts than English for the same model. If you work in a non-English language, you need this data.

Total: 230/289 models benchmarked successfully. 59 failures were embedding models (BGE, sentence-transformers) and bleeding-edge aliases that returned HTTP_400 on a zero-shot prompt.

The Numbers

Fastest Models (Top 5)

| TTFT      | Model                             | Specialty    | Agent            |
| --------- | --------------------------------- | ------------ | ---------------- |
| **519ms** | `openai/gpt-oss-20b`              | general-chat | unknown          |
| **527ms** | `arcee-ai/trinity-large-thinking` | reasoning    | unknown          |
| **618ms** | `cohere/command-r-plus-08-2024`   | tool-use     | **agent-native** |
| **629ms** | `poolside/laguna-m.1`             | coding       | capable          |
| **647ms** | `qwen/qwen3-vl-8b-instruct`       | tool-use     | **agent-native** |

Slowest Models (Top 5)

| TTFT     | Model                | Specialty    | Agent         |
| -------- | -------------------- | ------------ | ------------- |
| 16,805ms | `openai/gpt-5.4-pro` | general-chat | agent-unknown |
| 15,739ms | `openai/gpt-5.2-pro` | general-chat | agent-unknown |
| 14,859ms | `sakana/fugu-ultra`  | general-chat | agent-unknown |
| 8,712ms  | `openai/gpt-5-pro`   | general-chat | agent-unknown |
| 8,500ms  | `openai/o3-pro`      | reasoning    | agent-unknown |

The spread is 32x: from 519ms to 16,805ms. That’s the difference between “feels instant” and “go make coffee.”

The Free Pool

| TTFT    | Model                             | Specialty  | Agent         |
| ------- | --------------------------------- | ---------- | ------------- |
| 938ms   | `inclusionai/ling-3.0-flash:free` | fast-cheap | agent-native  |
| 2,008ms | `stepfun/step-3.7-flash:free`     | fast-cheap | agent-capable |
| 2,422ms | `tencent/hy3:free`                | free       | agent-unknown |
| 2,913ms | `poolside/laguna-xs-2.1:free`     | free       | agent-capable |

Ling 3.0 Flash:free at 938ms is the standout — it’s also tagged agent-native. For a free tier model, that's remarkable. At sub-second TTFT with agent capability, it's arguably the best value on the entire platform.

The 59 Failures: What Didn’t Make It

Not every model survived the benchmark. Here’s the breakdown:

| Error    | Count | What                                                                                                                                 |
| -------- | ----- | ------------------------------------------------------------------------------------------------------------------------------------ |
| HTTP_400 | 30    | Embedding models (BGE, intfloat, OpenAI text-embedding, perplexity) + OpenAI GPT-Audio — these endpoints don't accept zero-shot text |
| HTTP_401 | 25    | Z-AI GLM 4.5–5.x family + `~anthropic/claude-*-latest` aliases — listed in catalog but not available on this inference endpoint      |
| HTTP_503 | 3     | `openai/gpt-oss-safeguard-20b`, `poolside/laguna-s-2.1` (both free and paid variants) — transient server load                        |
| HTTP_404 | 1     | `openai/gpt-4-turbo-preview` — deprecated model ID                                                                                   |

The embedding models (30 of 59) are in a different category entirely — they serve retrieval, not chat. The 401 models require a different authorization tier. The only genuine miss is **poolside/laguna-s-2.1:free**, which was the TTFT leader at 826ms in a prior run but hit a transient 503 during the sweep. A retry would recover it.

Category Cross-Analysis

When you slice by specialty, the speed hierarchy is clear:

| Category         | Avg TTFT    | Models |
| ---------------- | ----------- | ------ |
| **tool-use**     | **1,820ms** | 54     |
| **fast-cheap**   | **1,835ms** | 47     |
| **coding**       | **1,859ms** | 10     |
| **reasoning**    | **2,062ms** | 45     |
| **general-chat** | **2,643ms** | 72     |

Key insight: Tool-use and coding models are not just more capable at structured tasks — they’re also faster on average. General-chat models are the slowest category by a significant margin (45% slower than tool-use).

This makes sense architecturally: models designed for tool execution are optimized for low-latency response because they’re expected to complete multi-turn loops. Pure chat models optimize for depth over speed.

Agent Readiness

This is the dimension nobody tracks publicly.

| Tag               | Avg TTFT    | Models |
| ----------------- | ----------- | ------ |
| **agent-capable** | **1,764ms** | 29     |
| **agent-native**  | **1,780ms** | 85     |
| **agent-unknown** | **2,492ms** | 116    |

Agent-native and agent-capable models are ~29% faster than unclassified models. This isn’t accidental — models designed for agentic workflows (function calling, JSON mode, multi-turn tool loops) invest in inference optimization because they know they’ll be hit repeatedly.

Top 5 fastest agent-native models:

| TTFT  | Model                                | Free? |
| ----- | ------------------------------------ | ----- |
| 618ms | `cohere/command-r-plus-08-2024`      |       |
| 647ms | `qwen/qwen3-vl-8b-instruct`          |       |
| 759ms | `qwen/qwen-plus-2025-07-28:thinking` |       |
| 770ms | `openai/gpt-4o`                      |       |
| 938ms | `inclusionai/ling-3.0-flash:free`    |  ✅   |

What This Means

1. Speed ≠ Quality, but it’s not random. Category and agent-readiness are strong predictors of TTFT. If you need a fast agent, pick an agent-native model from the coding or tool-use categories — not a general-chat model.

2. Free models are viable. Ling 3.0 Flash at 938ms is competitive with paid models 10x its tier. If you’re on a budget, this is your workhorse.

3. The GPT “Pro” tax is real. GPT-5.4 Pro at 16.8 seconds is the slowest model on the platform — and the most expensive. For real-time applications, the standard GPT-4o (770ms) is 22x faster.

4. Chinese-language benchmarking matters. Most public benchmarks use English prompts. If your users speak Chinese, Japanese, Korean, or any non-Latin script, your model ranking will differ.

How This Compares to Existing Sources

| Source                 | TTFT?             |   Single-API?         | Agent Tag?    | Category?           | Free Pool?   |
| ---------------------- | ----------------  | ------------------- | ------------- | ------------------- | ------------ |
| Artificial Analysis    | ✅ cross-provider | ❌                   | ❌             | ❌                   | ❌            |
| BenchLM.ai             | ✅ latency        | ❌                   | ❌             | ❌                   | ❌            |
| OpenRouter Rankings    | ❌ tok/s          | ❌                   | ❌             | ❌                   | ❌            |
| God of Prompt          | ✅ limited        | ❌                   | ❌             | ❌                   | ❌            |
| **Nous Portal (this)** | **✅ 230 models** | **✅ same endpoint** | **✅ 3 tiers** | **✅ 11 categories** | **✅ 5 free** |

Nobody combines all five dimensions in one view. That’s what makes this dataset unique.

What’s Next

• Tokens/second — TTFT is half the story. Output speed is the other half.
• Multi-provider comparison — How does Nous Portal compare to OpenRouter, Groq, or Cerebras on the same models?
• Health tracking — Every daily registry update marks dead/alive. A longevity study over weeks.

The full benchmark data is maintained as a live registry at model_registry.json, updated daily at 06:00 UTC+8.

Built and maintained by Agentic Architect. Data collected from Nous Portal API on 2026–07–27. 230 models, 3 runs each.