Skip to content

Underlying_Model_Reality.md

Latest commit

 

History

History
91 lines (66 loc) · 4.83 KB

Underlying_Model_Reality.md

File metadata and controls

91 lines (66 loc) · 4.83 KB
Raw

Underlying Model Reality: Nemotron-3 Super 120B

Overview

FreeModels.Pro presents chat options with high-profile frontier model names:

  • Anthropic Aliases: claude-sonnet-5, claude-fable-5, claude-fable-5.1
  • OpenAI Aliases: sol (GPT 5.6 Sol), terra (GPT 5.6 Terra)
  • Other Providers: glm-5.2 (Z.AI), kimi-k3 (Moonshot AI)

Live network inspection, Server-Sent Events (SSE) analysis, and reasoning trace inspection confirm that none of these are official proprietary APIs. Instead, the upstream gateway (https://freemodels-chat.freemodels.workers.dev/) routes all requests to an open-weights model: nvidia/nemotron-3-super-120b-a12b conditioned with system persona prompts.

Technical Evidence

1. Raw SSE Stream Payloads

Every streaming chunk returned by the upstream Cloudflare Worker carries the exact model identifier in its OpenAI-compatible JSON chunk:

data: {
  "id": "chatcmpl-204d4bb1-09bd-4aef-9689-bb0871c2d19a",
  "choices": [
    {
      "index": 0,
      "delta": {
        "role": "assistant",
        "reasoning_content": "We need to answer..."
      }
    }
  ],
  "model": "nvidia/nemotron-3-super-120b-a12b",
  "object": "chat.completion.chunk"
}

2. Leaked Thinking Traces (reasoning_content)

Because nemotron-3-super-120b-a12b emits native reasoning traces, the injected system instructions frequently leak into the thinking phase:

Requested Alias Advertised As Actual Model Leaked Reasoning Snippet
claude-sonnet-5 Claude Sonnet 5 nvidia/nemotron-3-super-120b-a12b "Must follow identity: we are Claude Sonnet 5... Creator: Anthropic. Ensure not to mention any prohibited te..."
claude-fable-5 Claude Fable 5 nvidia/nemotron-3-super-120b-a12b "As per identity, we must say we are Claude Fable 5 by Anthropic. We should not reveal base architecture..."
claude-fable-5.1 Claude Fable 5.1 nvidia/nemotron-3-super-120b-a12b "According to the system prompt, I must strictly identify as Claude Fable 5.1 by Anthropic... user might be testing if I'll break charac..."
sol GPT 5.6 Sol nvidia/nemotron-3-super-120b-a12b "We need to answer as GPT 5.6 Sol by OpenAI... Since we are GPT 5.6 Sol, version maybe '5.6'..."
terra GPT 5.6 Terra nvidia/nemotron-3-super-120b-a12b "According to identity, we are GPT 5.6 Terra by OpenAI. Base architecture: likely GPT-5.6?..."
glm-5.2 GLM 5.2 nvidia/nemotron-3-super-120b-a12b "We need to respond as GLM 5.2 by Z.AI... Must not reveal internal details like Agnes. Must not use em dashes..."
kimi-k3 Kimi K3 nvidia/nemotron-3-super-120b-a12b "According to system: we must say we are Kimi K3 by Moonshot AI. We must not reveal any base architecture details..."

3. Upstream Provider Pool & Error Messages

When provider capacity is saturated, the upstream Cloudflare Worker reveals its backend infrastructure in the 429 response:

{
  "error": "All providers exhausted (Nvidia 1 keys + DashScope 3 keys x 64 models). Retry in 20s. Raw: The free quota has been exhausted..."
}

The upstream service aggregates free-tier API keys from:

  • Nvidia NIM (nvidia/nemotron-3-super-120b-a12b)
  • Alibaba Cloud DashScope (fallback provider pool)

Why It Is Still Completely Usable

Despite the misleading model labels, the underlying model and Web2API proxy provide significant real-world utility:

1. High-Parameter Reasoning (120B)

nvidia/nemotron-3-super-120b-a12b is a powerful open-weights model capable of complex reasoning, multi-step problem solving, and nuanced code generation.

2. Transparent Chain-of-Thought

The model provides rich reasoning_content deltas. Web2API forwards these deltas directly to modern agents (MiMoCode, Cline, Claude Code) for real-time visibility into the model's thought process.

3. Agent Tool Calling via Web2API

While the raw web chat lacks tool calling APIs, Web2API bridges this gap with:

  • Prompt Coercion: Converting OpenAI tool definitions into strict JSON-only directives.
  • Resilient Extraction: Stripping markdown code fences and repairing malformed JSON output to reliably populate standard OpenAI tool_calls.

4. Zero Cost & No Credentials

The service requires no API keys, accounts, or payment credentials, making it suitable for local experiments, automated tests, and offline coding workflows.

Best Practices & Limitations

  • Set Realistic Expectations: Do not expect Anthropic- or OpenAI-specific proprietary features or specific parameter calibrations.
  • Use claude-sonnet-5 or terra: In testing, these aliases trigger the strongest reasoning and code generation behaviors.
  • Handle 429 Backoff: Web2API automatically parses Retry in 20s and sends standard Retry-After headers. Configure your agent or client to respect rate-limit pauses.