Qwen3.8-27B for hipfire

MQ4-quantized Qwen3.8-27B text tower for hipfire, the Rust-native LLM inference engine for AMD GPUs.

This artifact packages the 27B dense DeltaNet/attention language model in hipfire's native MQ4 container. It does not include the upstream vision encoder, so this repository supports text generation only.

Model

• Upstream: Qwen/Qwen3.8-27B
• Parameters: 27B
• Layers: 64
• Hidden dimension: 5120
• Vocabulary: 248,320
• Native context: 262,144 tokens
• Maximum output: 81,920 tokens
• Quantization: MQ4
• File: qwen3.8-27b.mq4
• Size: 14,980,361,216 bytes

Qwen3.8 thinking is enabled by default. Its reasoning_effort ladder is low, medium, high, and xhigh; hipfire passes the selected effort to the upstream chat template without rewriting it.

Usage

hipfire pull qwen3.8:27b
hipfire run qwen3.8:27b "Design a streaming Rust file deduplicator."

Hipfire's registry policy for this tag automatically selects Q8 KV, the VMM backend, the native 262,144-token context, and the 81,920-token output allowance. VMM reserves the logical context window while committing physical KV pages only as they are used.

Validation

The release artifact was validated on gfx1100 and gfx1201 through hipfire's native matrix benchmark and OpenAI-compatible serve path. Qwen3.8 stayed within 0.57% of Qwen3.6-27B across the tested prefill and decode matrix, and completed an eight-turn sampled coding/retrieval session with reasoning_effort=xhigh without empty output, runaway generation, token attractors, or retrieval misses.

License and attribution

The upstream Qwen3.8 weights are Apache-2.0 licensed. This quantized packaging was produced for hipfire; see the upstream model card and the hipfire repository for license, attribution, and provenance details.