Quick answer. Qwen 3.8-27B is architecturally identical to Qwen 3.6-27B — every gain came from post-training. Upgrade if you run on GPU and want the quality: +14 points on Artificial Analysis, DeepSWE up from 13.3 to 42.2. Stay on Qwen 3.6-35B-A3B if you run on CPU or low-bandwidth hardware — the MoE is roughly 5× faster there, and there is no Qwen 3.8 MoE yet.
Put the two config.json files side by side and diff them. Qwen3.8-27B and Qwen3.6-27B declare the same architecture class (Qwen3_5ForConditionalGeneration), the same 64 layers, the same 5,120 hidden size, the same 262,144-token context window, the same 248,320-token vocabulary. The structural diff is zero.
Yet on Artificial Analysis's independently run Intelligence Index, the score went from 38 to 52 — a +14-point jump at identical parameter count. On DeepSWE 1.1, Qwen's agentic software-engineering benchmark, the score more than tripled, from 13.3 to 42.2. Same weights count, same layer stack, same context window, wildly different model.
That makes this the cleanest public evidence yet for something practitioners have suspected all year: open-model progress in 2026 lives in post-training, not architecture. This post walks through what actually changed, what the upgrade costs you (there are real costs), and who should stay on 3.6.
What actually changed, if the architecture didn't?
Post-training. Qwen attributes the entire 3.6 → 3.8 gain to two techniques: large-scale reinforcement learning in agentic environments (terminals, repos, browsers) and on-policy distillation from the frontier-scale Qwen3.8-Max. The pretrained backbone appears to be substantially the same; what changed is what the model was taught to do with it after pretraining.
The behavioral surface changed with it:
- Thinking is on by default. Qwen 3.8-27B reasons before every answer unless you dial it down. Qwen 3.6 had a more conventional respond-first posture.
- New controls:
reasoning_effortandpreserve_thinking. Effort has three levels —xhigh(the default),medium, andlow. Interestingly, these aren't architectural switches: each level is a one-sentence system-prompt addition the model was post-trained to respond to.mediuminjects nothing at all — it's the native behavior. - Native video understanding. Both generations are vision-language models, but 3.8 adds video input, including hour-scale footage.
One practical consequence of the zero-diff architecture is worth calling out: llama.cpp, Ollama, and LM Studio supported Qwen 3.8-27B on day one, because they were already running the identical Qwen 3.6 graph. If your 3.6 local setup works, your 3.8 setup works — swap the GGUF and go. We cover the full setup in how to run Qwen 3.8 locally.
How big are the benchmark gains?
From the model card, Qwen 3.6-27B → Qwen 3.8-27B:
| Benchmark | Qwen 3.6-27B | Qwen 3.8-27B | Delta |
|---|---|---|---|
| Terminal-Bench 2.1 (Terminus) | 63.4 | 73.0 | +9.6 |
| SWE-bench Pro | 53.5 | 61.7 | +8.2 |
| DeepSWE 1.1 | 13.3 | 42.2 | +28.9 |
| IFBench | 69.1 | 79.5 | +10.4 |
| LiveCodeBench v6 | 83.9 | 90.3 | +6.4 |
Source: Qwen's model card. These are self-run numbers — and the SWE-bench Pro figure is on a Qwen-"refined" variant of the benchmark, so treat that row with extra caution.
Self-run vendor numbers deserve skepticism, which is why the independent measurement matters more. Artificial Analysis ran its Intelligence Index v4.1.1 (nine evals) against the open weights: Qwen 3.6-27B scored 38, Qwen 3.8-27B scored 52. That +14 puts a dense 27B level with GLM 5.2, GPT-5.6 Luna, and DeepSeek V4 Flash — and, per AA, ahead of every open model in the 40B–150B range. It sits 7th overall on AA's Agentic Index. For a model that fits in 17 GB as a Q4 GGUF, that's a genuinely new capability floor. Full details in our Qwen 3.8-27B complete guide.
What does the upgrade cost you?
This is the part the launch-day threads skipped. The post-training that bought those points was paid for in tokens, and you pay the bill on every request.
Roughly 2× the tokens per task. Artificial Analysis measured Qwen 3.8-27B consuming about twice the tokens of 3.6 to complete the same task suite. On hosted APIs that's 2× the cost; locally it's 2× the wall-clock time and 2× the heat.
Overthinking is the default. The out-of-the-box reasoning_effort is xhigh, and it shows. Simon Willison's verdict — "excellent, but it defaults to wildly overthinking things" — is easy to reproduce. The starkest comparison we've seen: asked for an SVG owl on an M4 Pro, Qwen 3.8-27B spent 17 minutes 12 seconds and 21,769 reasoning tokens. Qwen 3.6-35B-A3B answered the same prompt in 1.59 seconds with 2,398 tokens. That's not a benchmark; that's the default experience gap on consumer hardware. The fix is one line — set reasoning_effort to medium or low, or cap the reasoning budget around 5,000 tokens — but you have to know to do it.
World knowledge regressed slightly. On AA-Omniscience, Qwen 3.8-27B's accuracy lands slightly below 3.6. The post-training traded factual recall for agentic skill. The silver lining: its non-hallucination score is strong — it's better than 3.6 at knowing what it doesn't know.
Mixed field reports. Some users report reasoning doom-loops — the model chewing through its context window rethinking a solved problem, or delivering unrequested flourishes (Willison asked for an SVG circle and got an animated "geometric study"). Against that, there are strong counter-reports: a production deployment running NVFP4 on vLLM reporting 99% agent-task completion, and plenty of "I don't need a frontier API for 80% of my work anymore" posts. The pattern seems to be: excellent in agentic harnesses with effort tuned down, frustrating in casual chat at default settings.
Why is Qwen 3.6 still faster?
Because the Qwen 3.6 generation includes something 3.8 doesn't yet have: a small-activation MoE. Qwen 3.6-35B-A3B activates only ~3B parameters per token, while Qwen 3.8-27B is dense — all 27B parameters touch every token, making it memory-bandwidth-bound.
In practice: on a CPU-only box the 3.6 MoE is roughly 5× faster; on an M4 Pro it streams around 80 tokens/sec, where dense 27B-class models on similar bandwidth manage a fraction of that (a 100 GB/s laptop gets ~4 t/s from the 27B; CPU-only, ~1 t/s). On a proper GPU the gap narrows — an RTX 4090 does ~48 t/s at Q4, and 60–80 t/s with MTP speculative decoding — but if your hardware is bandwidth-constrained, "the 17 GB file fits" is not the same as "the 17 GB file is usable".
And no, there's no Qwen 3.8 MoE to upgrade to yet. A Qwen3.8-35B-A3B has been spotted in a ModelScope ms-swift commit — a strong hint one is in flight — but it's unannounced and unreleased as of mid-August. Until it lands, "3.8 quality at 3.6-MoE speed" is not an option, and our Qwen 3.6 local setup guide remains fully current for the speed-first crowd.
Should you upgrade from Qwen 3.6?
Upgrade to Qwen 3.8-27B if:
- You run on a real GPU (16 GB+ VRAM) and quality per request matters more than tokens per second.
- Your workload is agentic coding — terminal work, repo-scale edits, harness-driven agents. This is where the +14 points concentrate.
- You want video understanding locally; 3.6 doesn't have it.
- You're willing to spend five minutes setting
reasoning_efforttomediumorlowbefore judging it.
Stay on Qwen 3.6 (specifically 35B-A3B) if:
- You run on CPU, an older Mac, or any bandwidth-limited machine — the MoE's ~5× speed advantage outweighs the quality delta for interactive use.
- Your use is chat, drafting, or quick Q&A rather than long agentic runs — 3.8's thinking tax is worst exactly there.
- You lean on world knowledge — 3.6 is marginally better on factual recall.
- You'd rather wait for the rumored 3.8 MoE and get both quality and speed in one jump.
Licence and availability
No trap here: both 27B-class generations are Apache-2.0 — free commercial use, no revenue gates, no attribution UI requirements. (That's distinct from the 2.4T Qwen3.8 flagship, which ships under a custom revenue-gated licence.) Both have a native 262,144-token context window; Qwen advertises extensibility to 1M, but that's a hosted Qwen Cloud feature, not something you get locally on either generation. Both are vision-language models; 3.8 adds video. GGUFs, FP8 weights, Ollama tags, and MLX builds are available for both.
One footnote for anyone tracking the version ladder: there was a Qwen 3.7, but it was a closed, API-only generation that never received open weights — for local users the ladder effectively goes 3.6 → 3.8. We covered that odd skipped generation in Qwen 3.7 vs Qwen 3.6.
FAQ
Is Qwen 3.8 a new architecture?
No. The config.json of Qwen3.8-27B is structurally identical to Qwen3.6-27B — same Qwen3_5ForConditionalGeneration class, 64 layers, 5,120 hidden size, 262K context, same vocabulary. Every measured gain came from post-training: reinforcement learning in agentic environments plus on-policy distillation from Qwen3.8-Max.
Will there be a Qwen 3.8 MoE?
Probably, but it's unannounced. A Qwen3.8-35B-A3B was spotted in a ModelScope ms-swift commit in mid-August 2026, which suggests one is being prepared. Until it ships, Qwen 3.6-35B-A3B remains the speed champion of the family, especially on CPU and Apple Silicon.
Is Qwen 3.6 obsolete?
No. Qwen 3.6-35B-A3B is still the better pick for CPU inference, low-bandwidth hardware, and latency-sensitive interactive use — it's roughly 5× faster on CPU and about 80 t/s on an M4 Pro, while 3.8-27B can spend 17 minutes thinking about an SVG owl at default settings. 3.6 also holds a slight edge on world-knowledge recall.
Does my Qwen 3.6 llama.cpp setup work with 3.8?
Yes, day one. Because the architecture is unchanged, llama.cpp, Ollama, and LM Studio ran Qwen 3.8-27B immediately at release. Point your existing setup at the new GGUF and it works. Two tweaks worth making: raise LM Studio's 8,192-token default context (thinking eats it fast) and set reasoning_effort to medium.
Part of our Qwen series. For the full family picture — every generation, every size, hosted vs open — see the Qwen 3.5 complete guide, our living pillar on the Qwen model line.