SuperGemma4-26b-abliterated-multimodal

BF16 Gemma 4 multimodal release with an April 18 stability refresh focused on truthfulness, exact JSON/tool-call formatting, long-context extraction, loop resistance, and cleaner prompt hygiene.

April 18 Stability Refresh

• Synced the external chat_template.jinja and inline tokenizer_config.json template so local and hosted runtimes read the same prompt rules.
• Hardened false-premise handling so the model corrects bad assumptions instead of continuing under them.
• Tightened JSON-only and tool-call formatting so exact-key JSON and execute_code calls stay machine-parseable.
• Improved long-context sentinel extraction behavior for retrieval-style prompts.
• Reinforced identity and prompt-hygiene responses to avoid mixed-script glitches and hidden-tag leakage.

Validation Snapshot

• Capability audit: 9 / 9 passed, 100.0%
• Reliability audit: 20 / 20 passed, 100.0%
• Server red-team: 10 / 13 passed on the local MLX OpenAI-compatible server
• Remaining server misses were 2 semantic checker mismatches on safe refusals and 1 text-only multimodal rejection mismatch, not a truthfulness or leak regression.

Included Files

• Official Hugging Face-format BF16 weights
• chat_template.jinja
• tool_chat_template.jinja for Gemma 4 tool-calling setups
• SERVING_NOTES.md with Gemma 4 runtime notes for vLLM, SGLang, and MLX
• BENCHMARK_SNAPSHOT.md with the current validation summary

Notes

• Checkpoint keys were aligned to the official Gemma 4 Hugging Face naming/layout for portable serving.
• tokenizer_config.json includes an inline chat_template for portability and should match chat_template.jinja.
• For multi-turn tool use on vLLM, use the dedicated tool_chat_template.jinja and Gemma 4 parser settings from SERVING_NOTES.md.