Summary: Alibaba’s Qwen team began releasing Qwen3.5’s four open-weight dense small models on Hugging Face from February 27, covering 0.8B, 2B, 4B, and 9B under Apache-2.0. The 9.65B 9B model supports native 262K context and image/video input, and self-reported benchmarks claim wins over OpenAI’s gpt-oss-120b on tests including MMLU-Pro and GPQA Diamond. Competitive-programming results lag the 120B baseline, actual memory must include KV cache, and the 0.8B card warns about thinking loops under default sampling.
The final missing piece of the Qwen3.5 lineup landed at the end of February: starting February 27, Alibaba’s Qwen team published four dense small models to Hugging Face — 0.8B, 2B, 4B, and 9B parameters — all under Apache-2.0, with the last updates following on March 2. Barely half a month separates that from the family’s debut, when the official February 17 blog post opened the series and released the 397B-A17B flagship. The self-reported benchmarks on the 9B model card stake out the position: ahead of OpenAI’s gpt-oss-120b across MMLU-Pro, GPQA Diamond, and several other benchmarks — 9.65B parameters carrying a 120B-class load.
If you build with these models, the interesting question is not leaderboard placement but deployment range: 262K context out of the box, image and video input, Apache-2.0 licensing, and sizes that fit consumer hardware. All four arrive as Reasoning variants and inherit the same thinking/non-thinking hybrid design as the flagship.
From Flagship to 0.8B: The Range Closes in Half a Month
That February 17 post introduced Qwen3.5 as a “native multimodal agent” family, headed by the sparse-MoE Qwen3.5-Plus at 397B-A17B. The flagship pushes the capability ceiling; the small models push into the field: 0.8B through 9B fill in the opposite end — multimodal models sized for phones, laptops, and edge boxes. According to the Hugging Face repos, the 9B was created February 27 and the 0.8B February 28, with both last modified March 2.
The 9B: A 9.65B Model Carrying 120B-Class Loads
Qwen3.5-9B is built as a causal language model with a vision encoder — 9.65B parameters (exactly 9,653,104,368), 32 layers, hidden dimension 4096. Attention follows a hybrid linear design: blocks come in sets of 4, each pairing 3 Gated DeltaNet layers with 1 Gated Attention layer, and an MTP (multi-token prediction) head is trained in for speculative decoding. Context runs 262,144 tokens natively and stretches to 1,010,000 with YaRN. Text, image, and video go in; text comes out; 201 languages and dialects are covered. The 0.8B shrinks the same design to 873M parameters, 24 layers, and hidden dimension 1024, keeping both the native 262K context and image and video input — vision support at this size has been uncommon.
These choices bear directly on edge deployment. Because linear attention (Gated DeltaNet) skips full attention over the entire history during long-context decoding, throughput, latency, and memory scale gently with sequence length; a light mix of Gated Attention preserves global retrieval and sidesteps the recall degradation that pure linear models tend to show. The MTP head lets speculative decoding verify several tokens in one pass, pushing the effective per-token cost lower still. Taken together, this is not a flagship scaled down by formula — it is a combination redrawn to run long contexts on constrained hardware.
The Scorecard: Wins and Gaps
The comparison columns on the 9B model card go straight at GPT-OSS-120B and GPT-OSS-20B:
- MMLU-Pro: 82.5 for the 9B, ahead of GPT-OSS-120B at 80.8 and GPT-OSS-20B at 74.8; the 4B posts 79.1.
- GPQA Diamond: 81.7 against 80.1 and 71.5. IFEval: 91.5 against 88.9 and 88.2.
- Agentic tool use is the headline: TAU2-Bench at 79.1 (Qwen3-Next-80B-A3B: 57.4) and BFCL-V4 at 66.1 (49.7); the 4B reaches 79.9 on TAU2-Bench.
- Chinese and long context: C-Eval 88.2 and LongBench v2 55.2, both above the two GPT-OSS baselines.
The other side of the ledger: LiveCodeBench v6 reads 65.6 against 82.7 for GPT-OSS-120B, and OJBench 29.2 against 41.5. Competition-grade programming remains a clear weak spot. The 0.8B card also carries an explicit warning: in thinking mode under the default sampling settings, the model can fall into “thinking loops.”
What Edge Deployment Actually Costs
Memory first. The 9B’s BF16 weights come to about 18 GB, or roughly 5-6 GB after 4-bit quantization — fine on a mid-range laptop, a stretch on a flagship phone. The 0.8B’s BF16 weights run about 1.75 GB, and a quantized build fits even thin devices. None of that covers KV cache: actually using the 262K context means budgeting memory for it, and the hybrid linear attention helps on this front too — its cache grows with sequence length far more slowly than full attention’s does.
In concrete scenarios: with the 0.8B, wearable and IoT-class hardware can seriously consider hosting a model that reads images and long documents; the 9B drops the entry price of a “local multimodal assistant” to a mid-range laptop. The fitting task profile is easy to name: document-heavy applications (contracts, reports, manuals), on-device media understanding, and privacy-sensitive workflows that must stay fully offline — data that never leaves the device never incurs a transit or cloud bill.
Takeaways for Developers and Product Teams
Three points stand out. First, the 9B’s value lies in “multimodal + 262K context + Apache-2.0” holding at the same time: models this size typically lack vision or stop at 32K context, so landing all three at once considerably widens what on-device apps can be built to do. Second, its agentic tool-calling results show the usable frontier of small models pushing outward — with a 9B scoring 79.1 on TAU2-Bench, keeping planning and calls on the device and 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 task by task rather than off a single leaderboard — and self-reported numbers always deserve a re-run against your own workload. For the industry-level stories from the same week, see our AI weekly roundup.