TaimoorSiddiqui

HopCoder-Mini-35B-A3B-VL36

README

License: apache-2.0

Identity

Product name: HopCoder Mini
Repository: TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36
Creator and publisher: Taimoor Siddiqui
Default chat identity: injected by chat_template.jinja

The chat template includes a default system identity so normal chat, vision, and tool-calling prompts identify the assistant as HopCoder Mini by Taimoor Siddiqui. Runtime architecture, tokenizer, and processor class names in JSON config files are intentionally left unchanged when required by Transformers or vLLM loaders. Do not rename those loader identifiers unless the corresponding runtime classes are also renamed and registered.

What Changed

Added the VL36 model.visual.* tensors to the agentic text checkpoint.
Removed the language_model_only flag so the vision tower is instantiated.
Added HopCoder Mini metadata to config.json.
Added an identity-safe default system instruction to chat_template.jinja.
Added Modal-only repo and benchmark runners in modal_repo_ops.py and modal_h200_eval.py.
Added a smallest-useful-first LoRA SFT path in modal_small_sft.py.
Added tools/identity_scan.py for repository-level identity leak checks.

Result: 1026 tensors total, with the text backbone and lm_head preserved and the vision tower loaded from model-vision.safetensors.

Serving

Use vLLM 0.23 or newer for the later benchmark/serving run.

bash
vllm serve TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36 \
  --tensor-parallel-size 1 \
  --limit-mm-per-prompt '{"image": 4, "video": 1}'

Do not pass --language-model-only; that disables the vision tower.

Capabilities

Agentic coding and tool-calling workflows.
Image and video input through the VL36 vision stack.
Long-context chat and coding use cases supported by the checkpoint config.
BF16 full-precision weights; this is not a quantized release.

Known Limitations

Vision quality is not benchmarked. The VL36 vision tower is a cross-generation graft from a separate checkpoint and may differ from a natively co-trained VLM. No vision benchmark suite has been run.
No additional safety tuning or evaluation beyond basic smoke tests. Identity probes and a short vision smoke test passed, but no formal safety, alignment, or red-team evaluation has been conducted.
Agent-style interactive coding is limited. The base backbone was not trained on interactive shell-agent workflows; the model can read and reason about code but struggles to transition from exploration to autonomous file editing in an unsupervised agent loop.

Current Evaluation Status

This preparation run does not execute the vLLM benchmark. The H200 Modal runner is included so the benchmark can be launched later with the requested resources: 8 vCPU, 32 GiB RAM, and an H200 GPU.

Prepared entrypoints:

bash
python -m modal run modal_repo_ops.py --action repo-scan
python -m modal run modal_h200_eval.py --action identity-probe --model-id TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36
python -m modal run modal_h200_eval.py --action vision-smoke --model-id TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36
python -m modal run modal_h200_eval.py --action suite-command --suite swebench-pro --command "<official SWE-bench Pro command>"
python -m modal run modal_h200_eval.py --action suite-command --suite terminal-bench-2 --command "<official Terminal-Bench 2.0 command>"
python -m modal run modal_h200_eval.py --action suite-command --suite claw-eval --command "<official Claw-Eval command>"

See BENCHMARKS.md for the deferred benchmark workflow.

See TRAINING.md for the smallest useful first LoRA SFT pass and the Stage 2 reasoning/code/agent trace-mix plan.

Rebuild

The included build_merged_vl36.py script can reproduce the local merged checkpoint from the source text and vision repositories. It is intended for controlled rebuilds, not for routine inference.

bash
pip install huggingface_hub safetensors torch
python build_merged_vl36.py

License

Released under Apache-2.0. See LICENSE and NOTICE for copyright and third-party license notices retained for compliance.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

TaimoorSiddiqui

Model Tree

Base

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality