Qwen3.5-27B vs Qwen3.6-27B

SAVRN's Notes on Qwen3.5-27B

27.8B parameters in a dense multimodal body, released February 2026: it reads images and text over a 262,144-token context. Memory runs 66.7 GB at 16-bit, 33.3 GB at 8-bit and 16.7 GB at 4-bit — the cheapest listed setup is one 192 GB MI300X at $1.85 an hour, and at 4-bit a single workstation card holds it. The repo is 11 safetensors shards totalling 55.6 GB, and downloads run 1.9M a month.

Apache 2.0 permits commercial use with notices kept. The card states compatibility with Transformers, vLLM, SGLang and KTransformers. Reported results on the page include SWE-bench 72.4, MMLU-Pro 86.1, Terminal-Bench 2.0 41.6 and ScreenSpot-Pro 70.3, plus HLE at 48.5 with tools against 24.3 chain-of-thought — a gap that says tool access is not optional for this model's harder tasks. None of these is a SAVRN measurement.

If you would rather rent by the token, the Index observes DeepInfra at $0.26 in and $2.60 out per million tokens and Novita at $0.30 and $2.40, both read Oct 1, 2026. Against $1.85 an hour for self-hosting, hosted inference wins at low volume and loses quickly at sustained load — run the arithmetic at your actual traffic before committing to a card.

SAVRN Research

SAVRN's Notes on Qwen3.6-27B

Two Index hosts sell this one by the token, so the rent-or-own math is explicit: DeepInfra at $0.32 in and $3.20 out per million tokens, OVHcloud at $0.47 and $3.19, against one MI300X at $1.85 an hour that holds the full 16-bit working set of 66.7 GB — or 16.7 GB at 4-bit, leaving most of the 192 GB for the 262,144-token context. Third-party-reported evals top its MoE sibling on this list: 87.8 GPQA-diamond, 94.1 AIME 2026, 77.2 SWE-bench Verified.

Apache 2.0 permits commercial use, modification and redistribution, notices kept and changes stated, so a fine-tune is yours to ship. One naming trap: the configuration architecture reads Qwen3_5ForConditionalGeneration with model type qwen3_5 despite the 3.6 name — confirm your serving stack loads it before committing. No base model or paper is on record. Dense 27.8B parameters, image-and-text in, text out, released April 2026.

SAVRN Research