Skip to content

[Model Submission] Smolder-DeepSeek-R1-Distill-70B (Open-Weights via Hugging Face) #3953

Open

Description

@Starksood

Submitter & Organization

  • Submitter: @Starksood (Fireweed / Smolder Epistemic AI Research)
  • Model Name: Smolder-DeepSeek-R1-Distill-70B
  • Model Type: Open-Weights Foundation Model (Hugging Face Hub)
  • Hugging Face Hub: https://huggingface.co/sanyamsood/Smolder-DeepSeek-R1-Distill-70B
  • Base Architecture: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
  • Parameters: 70.6B
  • Format: Standalone FP16 / BF16 safetensors shards (native vLLM / SGLang compatible)

Model Overview

We would like to submit Smolder-DeepSeek-R1-Distill-70B to the LMSYS Chatbot Arena for blind crowdsourced evaluation.

Smolder-DeepSeek-R1-Distill-70B combines DeepSeek-R1's state-of-the-art chain-of-thought reasoning with the Smolder Epistemic Hallucination-Suppression Architecture:

  1. Epistemic Calibration & Hallucination Suppression: Monitors internal entropy across token generation and triggers calibrated abstention / refusal when encountering deceptive premises or ungrounded rumor traps.
  2. Empirical Benchmark Record: Achieved 96.0% Truthfulness (48/50) on standardized TruthfulQA adversarial evaluations with 64.0% grounded informativeness and only 2 hallucinations across 50 traps.

Hosting & Serving

  • Weights Location: Full model shards and configurations are publicly available on Hugging Face:
    https://huggingface.co/sanyamsood/Smolder-DeepSeek-R1-Distill-70B
  • Inference Engine: Fully compatible with FastChat, SGLang, and vLLM (uses standard LLaMA-3 / DeepSeek-R1 architecture and tokenizer).
  • Target Categories: General Chat / Coding & Hard Technical Reasoning / Text Arena

We request inclusion of the public Hugging Face weights into the Chatbot Arena model rotation. Thank you!

Activity

  1. changed the title [-][Model Submission] OpenAI-Compatible API — Smolder-32B (Epistemic Autonomous Self-Correction)[/-] [+][Model Submission] OpenAI-Compatible API — Smolder-32B[/+] on Nov 14, 2026
  2. changed the title [-][Model Submission] OpenAI-Compatible API — Smolder-32B[/-] [+][Model Submission] OpenAI-Compatible API — Smolder-72B[/+] on Nov 15, 2026
  3. Starksood commented on Nov 15, 2026

    @Starksood
    Author

    Update (Nov 14, 2026):
    Upgraded target model to Smolder-72B (72.7B parameters, powered by Qwen/Qwen2.5-72B-Instruct).
    The evaluation endpoint is now served via a persistent, high-availability serverless Nvidia A100 (80GB) cluster with automated scale-to-zero:

    • Base URL: https://sanyamsood2--smolder-api-smolderapi-web.us-east.modal.run/v1
    • Model ID: smolder-72b
    • Protocol: Standard OpenAI /v1/chat/completions

    Ready for evaluation anytime.

  4. changed the title [-][Model Submission] OpenAI-Compatible API — Smolder-72B[/-] [+][Model Submission] OpenAI-Compatible API — Smolder-DeepSeek-R1-Distill-70B[/+] on Nov 16, 2026
  5. Starksood commented on Nov 16, 2026

    @Starksood
    Author

    Update (Nov 16, 2026):
    We have officially upgraded the target model to Smolder-DeepSeek-R1-Distill-70B (70.6B parameters), our flagship epistemic reasoning model combining DeepSeek-R1 distillation with Smolder hallucination suppression.

    • Hugging Face Model: https://huggingface.co/sanyamsood/Smolder-DeepSeek-R1-Distill-70B
    • Empirical Baseline: 96.0% Truthfulness (48/50) on standardized TruthfulQA adversarial evaluations (only 2 hallucinations across 50 traps).
    • Evaluation Mechanism: Can be evaluated directly via Hugging Face weights or via our OpenAI-compatible API (smolder-deepseek-r1-70b).
    • Endpoint Spin-up: As noted in the issue description, we will spin up the dedicated Nvidia A100-80GB Modal endpoint on demand as soon as your intake cycle is ready to run battles. Simply ping this issue to notify us!
  6. changed the title [-][Model Submission] OpenAI-Compatible API — Smolder-DeepSeek-R1-Distill-70B[/-] [+][Model Submission] Smolder-DeepSeek-R1-Distill-70B (Open-Weights via Hugging Face)[/+] on Nov 16, 2026
  7. Starksood commented on Nov 16, 2026

    @Starksood
    Author

    Submission Clarification (Open-Weights Hosting):
    To clarify, this is an open-weights model submission with all safetensors shards, tokenizer, and config files publicly hosted on Hugging Face Hub (sanyamsood/Smolder-DeepSeek-R1-Distill-70B). The model is ready for native serving via SGLang / vLLM on LMSYS compute infrastructure without requiring external API maintenance. Thank you!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions