Qwen3.8-27B is now live at Agnify.ai

Qwen3.8-27B weights are released and this model is already fully supported in Agnify. Let's take a quick look at the release, why everyone is excited about this model and what you can do with it in our platform.

TL;DR: Qwen3.8-27B is now the default vision-language model on Agnify, and pipelines pinned to the old Qwen3.6-35B-A3B id upgraded themselves. The median gain across the vision benchmarks is +24%, but it is lopsided - agentic and document tasks jump, pure perception moves 2-3 points. It is also a dense model replacing a 3B-active MoE, so decode runs at roughly 40% of the old speed while time to first token barely moves.

Benchmarks

Qwen3.8 is a multimodal model, meaning it can take text, images or video as an input, reason on them and return a text output. And it actually is pretty capable. Look at the models it is being compared to:

Text & agentic performance

Vision-language performance

Bold marks the best score in each row. Where the model card reports both settings, the score in parentheses is with CI prompting enabled, and with-CI scores are only ever compared against other with-CI scores. A dash (-) means the result was not reported for that model.

Yeah, one of them is Opus 4.6 Max, which just a few months ago was practically state of the art. Not saying the 27B model is smarter overall, but in some areas - yes. Of course we are more interested in the visual capabilities, which also had a nice boost from the predecessor. So this update brings us +24% across all visual benchmarks - the median of the per-benchmark gains over the Qwen 3.6 generation, from +2% on OmniDocBench to +127% on BabyVision.

One thing worth being honest about: that gain is lopsided. The biggest jumps are agentic and computer-use benchmarks - OSWorld +20 points, Vision2Web +18, RecreationBench +17. The two pure perception benchmarks we look at below, RealWorldQA and ERQA, move 2 and 3 points. Both facts are in the same table.

Note that we upgraded from Qwen3.6-35B-A3B, which had a little weaker performance as it is a Mixture of Experts model, although it is faster. We benchmarked the latency on our inference hardware, and the trade-off is uneven. Time to first token - what you actually feel on a short visual answer - barely moved: ~25% on average, 0 to +0.3 s in absolute terms, and flat at 0.16 s for a single image at concurrency 1. Decode is where you pay. The dense 27B generates at roughly 40% of the MoE's speed (245 → 99 tok/s on a single image), which pushes total p50 to ~2.7× and cuts throughput to about 40% of what we had. We run with thinking off, so this is not extra reasoning time - it is the cost of activating 27B parameters per token instead of 3B.

Let's look into two visual benchmarks to gain some intuition on the types of things they test models for.

RealWorldQA

Real-world photos with questions about spatial layout, traffic rules and physical intuition - the kind of judgment a model needs to be useful in the real world.

Q: Can we make a left turn? Please answer directly with a single word or number.

Q: Where is the biggest bike? A. The red bike in the middle is the biggest bike. B. The red bike on the far left is the biggest. C. The black bike on the far right is the biggest. Please answer directly with only the letter of the correct option and nothing else.

ERQA

Embodied reasoning - questions about everyday scenes an agent has to act in: spacing, trajectories, occlusion and physical constraints.

Q: Where should another hanger be placed so that all hangers are spaced apart roughly equally? Choices: A. Purple. B. Blue. C. Red. D. Green. Please answer directly with only the letter of the correct option and nothing else.

Q: Which colored trajectory should be followed to cover the window with the more opaque of the two curtains? Choices: A. Purple. B. Green. C. Orange. D. Red. Please answer directly with only the letter of the correct option and nothing else.

None of these are hard for a person. All four are the kind of question that used to need a much bigger model.

Usage in Agnify

Qwen3.8-27B is now the default vision-language model across the platform - and pipelines pinned to the old Qwen3.6-35B-A3B id resolve to it automatically, so nothing needed re-wiring. Here is what you can point it at.

• Four nodes call it directly:
• Frame Description - ask a question about a frame, get an answer back.
• Attribute Extractor - detect and track objects, then extract preconfigured attributes for each one once its track ends.
• Object Inspector - one call per newly-confirmed object: describe it once, or verify it to get rid of detector false positives.
• State Classifier - sort each frame, or each zone you draw, into one of your own categories.
• Better video understanding - temporal mode sends a real, timestamped clip (up to 32 frames) instead of loose stills, so motion questions a single frame can't answer land: "is the person falling?", "which way is the forklift moving?".
• Reference images - upload labelled example images that ride along with every call, so the question becomes "which of these labelled pictures does this look like?" rather than "which prose description fits?".
• Structured output - constrain the answer to a JSON schema you define, and get back validated, typed fields instead of prose your next node has to parse.
• Self-hosted on our own H200, FP8 end to end - official FP8 checkpoint and FP8 KV cache via vLLM.

Check it out at app.agnify.ai.