Qwen3.8-27B

27B dense with a vision encoder, built on the Qwen 3.5 architectural foundation. Hybrid attention layout - Gated DeltaNet (linear attention) layers interleaved with Gated Attention in a 3:1 pattern across 64 layers - plus a ~248K vocabulary and Multi-Token Prediction (MTP) training. Multimodal: text, images, and hour-scale video in; text out.

• Context: 262,144 tokens native, extensible to ~1M via YaRN.
• Reasoning and tools: hybrid thinking - on by default, disable per request; reasoning_effort (xhigh default / medium / low) and preserve_thinking. Improved tool calling, including parsing of nested JSON objects, plus Developer Role support for agentic tools.
• Family: the open-weight sibling of cloud-only Qwen3.8-Max; also ships alongside the open Qwen3.8-2.4T-A95B (thinking-only).

Open weights under Apache-2.0 at Qwen/Qwen3.8-27B - released 2026-08-05, with a v2 weights update on 2026-08-14. Unsloth had day-zero quant access: its Qwen3.8 quants passed 5.1M downloads in the first five days.

Unsloth Dynamic 3.0 (2026-08-19). The re-quant calibrates on a higher-quality imatrix dataset (agentic coding, chat, multilingual) and improves layer selection - Unsloth measures >10% better top-1 accuracy at the same file size than every other quant provider (Divergence-300 @32, KL Divergence). The MTP module is dropped from UD-Q2_K_XL and below, saving ~500MB. What fits where (sizes are exact HF file sizes):

• UD-Q4_K_XL (17.56GB): the quality pick - ~19GB RAM, so 24GB-class VRAM (RTX 4090/5080) or any 24GB+ Mac.
• UD-Q2_K_XL (9.83GB): the Dynamic 3.0 headline quant - +8% top-1 accuracy vs the next-best provider at the same size. ~11GB RAM.
• UD-IQ1_S (6.19GB): 1-bit, 89% smaller than BF16, runs in 8GB RAM at ~72-77% top-1 agreement (72% per the docs, 77% in the launch announcement).
• NVFP4: vLLM on NVIDIA Blackwell only - ~1.5x faster than BF16 with 92-97% accuracy recovery, fits 24GB VRAM.

Run the GGUFs via llama.cpp -hf or Unsloth Desktop; there is no verified local Ollama tag yet. Honest framing: “by far the strongest model for its size” is Unsloth’s characterization - Model2-Max 3B still beats it on ARC-Challenge and MMLU-Pro - and the Dynamic 3.0 deltas are vendor-reported until independently replicated.

AMD Day 0 local support (2026-08-14). AMD validated the model on launch day for Windows + llama.cpp (Vulkan backend), publishing preliminary token-generation throughput averaged over 3+ runs:

• Up to 24.5 t/s on Ryzen AI Max+ 395 (MTP=4) - GMKtec EVO-X2 mini PC, 128GB unified memory with VGM set to 64GB, Adrenalin 26.7.1.
• Up to 51.8 t/s on a single Radeon AI PRO R9700 32GB (MTP=2) - Ryzen 9 9950X host with 64GB system RAM.
• ~24GB of variable graphics memory or VRAM to run comfortably - in line with the UD-Q4_K_XL recommendation above; older supported AMD platforms with 24GB+ also work.

Fastest paths on AMD hardware: LM Studio for a no-code start - set MTP draft tokens to 4 on Ryzen AI Max+ or 2 on the R9700, and uncheck “Try mmap” in the advanced model load settings - or the Lemonade local inference layer when you are wiring the model into an application. AMD calls the numbers preliminary and expects them to improve as Day 0 support matures. Source: AMD blog, Aug 14 2026.

Agents on Rails benchmark (Aug 2026, Le Mans round). 76.2% accuracy on 63 runs - the strongest open-weight model in the 16-model field, matching cloud-only Muse Spark 1.2 and ahead of GPT-5.6 Luna (73%) and Gemini 3.7 Flash (71.4%). The local-vs-hosted story: a 27B model you can run on your own hardware lands mid-pack against frontier cloud models. The catch is time - 27m 04s median (1.7x the next-slowest) - and Rails API recall is the worst in the field at 7.9%, so it often hand-rolls fixes instead of reaching for the right API.