Starting February 27, Alibaba’s Qwen team put the final piece of the Qwen3.5 lineup on Hugging Face: four dense small models at 0.8B, 2B, 4B, and 9B parameters, Apache-2.0 licensed, with the last updates landing March 2. That is barely half a month after the family opened with the 397B-A17B flagship announced in the official February 17 blog post. The self-reported benchmarks on the 9B model card set the positioning: it beats OpenAI’s gpt-oss-120b on MMLU-Pro, GPQA Diamond, and several other measures — 9.65B parameters doing 120B-class work.
For developers, the point is not the leaderboard but the deployment spectrum: native 262K context, image and video input, Apache-2.0 licensing, and sizes small enough for consumer hardware. All four ship as Reasoning variants sharing the same thinking/non-thinking hybrid design as the flagship.
From 397B to 0.8B: Filling Out the Lineup in Half a Month
The February 17 announcement framed Qwen3.5 as a “native multimodal agent” family, led by the sparse-MoE Qwen3.5-Plus at 397B-A17B. The flagship chases the ceiling; the small models chase deployment: 0.8B through 9B cover the other end — multimodal models that run on phones, laptops, and edge boxes. The Hugging Face repos show the 9B created February 27 and the 0.8B created February 28, both last modified March 2.
The 9B: Doing 120B-Class Work With 9.65B Parameters
Qwen3.5-9B is a causal language model with a vision encoder: 9.65B parameters (9,653,104,368), 32 layers, hidden dimension 4096. The architecture is hybrid linear attention — in each group of 4 blocks, 3 Gated DeltaNet layers pair with 1 Gated Attention layer — plus an MTP (multi-token prediction) head trained for speculative decoding. Native context is 262,144 tokens, extensible to 1,010,000 via YaRN. It takes text, image, and video input, outputs text, and covers 201 languages and dialects. The 0.8B is the same architecture shrunk: 873M parameters, 24 layers, hidden dimension 1024, the same native 262K context and the same image and video input — vision support at this size has been rare.
The architecture choices matter directly for edge deployment. Linear attention (Gated DeltaNet) means long-context decoding does not run full attention over the entire history, so throughput, latency, and memory grow gently with sequence length; mixing in a few Gated Attention layers preserves global retrieval and avoids the recall degradation pure linear models often show. The MTP head lets speculative decoding verify multiple tokens at once, further cutting effective per-token cost. In other words, this is not a big model scaled down by rote — it is a combination redesigned to run long contexts on constrained hardware.
Benchmarks: Where It Wins and Where It Loses
The 9B model card names GPT-OSS-120B and GPT-OSS-20B as the comparison columns:
- MMLU-Pro: the 9B scores 82.5, above GPT-OSS-120B at 80.8 and GPT-OSS-20B at 74.8; the 4B reaches 79.1.
- GPQA Diamond: 81.7 versus 80.1 and 71.5. IFEval: 91.5 versus 88.9 and 88.2.
- Agentic tool use is the standout: TAU2-Bench at 79.1 (Qwen3-Next-80B-A3B: 57.4) and BFCL-V4 at 66.1 (49.7). The 4B even posts 79.9 on TAU2-Bench.
- Chinese and long-context work: C-Eval 88.2 and LongBench v2 55.2, both above the two GPT-OSS baselines.
The honest side: LiveCodeBench v6 sits at 65.6 against 82.7 for GPT-OSS-120B, and OJBench at 29.2 versus 41.5. Competitive programming remains a clear weak spot. The 0.8B card also carries a specific warning: in thinking mode with default sampling settings, the model is prone to “thinking loops.”
The Realities of Edge Deployment
Run the memory math. The 9B’s BF16 weights come to roughly 18 GB, or about 5-6 GB at 4-bit — workable on a mid-range laptop, marginal on a flagship phone. The 0.8B’s BF16 weights are about 1.75 GB, and quantized versions fit even thin devices. None of that includes KV cache: to actually use the 262K context, you need to budget for it, and the hybrid linear attention helps here too — the cache grows far more slowly with length than full attention’s does.
Translated into scenarios: the 0.8B makes a model that can read images and long documents plausible even on wearable and IoT-class hardware, while the 9B pushes the entry point for a “local multimodal assistant” down to a mid-range laptop. The task profile that fits is clear: document-heavy applications (contracts, reports, manuals), on-device media understanding, and privacy-sensitive workflows that must run fully offline — data that never leaves the device has no transit or cloud bill.
What It Means for Developers and Product Teams
Three takeaways. First, the 9B’s value is that “multimodal + 262K context + Apache-2.0” all hold at once: models at this size usually lack vision or stop at 32K context, and getting all three together noticeably widens the design space for on-device apps. Second, its agentic tool-calling scores show the usable boundary of small models pushing outward — when a 9B posts 79.1 on TAU2-Bench, keeping planning and calls on the device while sending only the necessary steps to the cloud becomes a workable cost architecture. Third, do not read only the wins: the competitive-programming gap is a reminder to pick models per task, not per single leaderboard — and self-reported numbers always deserve a re-run on your own workload. For the industry-level stories of the same week, see our AI weekly roundup.
Sources
- Qwen3.5-9B — Hugging Face
- Qwen3.5 Collection — Hugging Face
- Qwen3.5: Towards Native Multimodal Agents — Alibaba Cloud Blog
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.