Alibaba releases weights for Qwen3.8 models under Apache 2.0 license, including Qwen3.8-27B, which it says beats Qwen3.7-Plus and excels in real-world coding
We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With just 27B parameters, it outperforms Qwen3.7-Plus overall and shines in real-world coding & office workflows. - 262K native context, easily extendable to 1M
@alibaba_qwen
Context & Ripple Effects
Alibaba’s Qwen line has already paired large-scale open-weight releases with smaller variants: February’s 397B-parameter multimodal Qwen3.5 release was followed by an April 27B dense model that Alibaba said surpassed a much larger predecessor on coding benchmarks. Qwen3.8 extends that cadence with another 27B-class release aimed at coding and office use.
The key continuity is distribution as much as model size: Alibaba is making a newer native multimodal model available as weights under Apache 2.0, rather than limiting the advance to a hosted Qwen offering.
First-order effects
Developers and organizations that use Qwen can evaluate and deploy Qwen3.8-27B’s released weights, including its 262K native context window, in place of older Qwen options.
Alibaba positions Qwen3.8-27B directly against Qwen3.7-Plus, making its claimed coding and office-workflow gains the immediate basis for migration decisions within the Qwen ecosystem.
Second-order effects
Alibaba’s earlier open-weight 27B dense model now has a newer successor to benchmark against, increasing pressure on Qwen’s own model lineup to differentiate by workflow performance rather than parameter count alone.
Tooling providers and deployers serving Qwen users gain another multimodal, long-context target, encouraging support to follow the released-weight model family rather than a single hosted endpoint.
Third-order effects
If Alibaba continues pairing frequent weight releases with claims that compact dense models match or exceed larger predecessors, model buyers gain more leverage to choose models by deployment fit and task performance instead of scale alone.
The pattern points toward open weights functioning as an enduring distribution channel for Qwen, with value shifting toward deployment, integration, and workflow-specific evaluation around the models.
The trend: Alibaba is using successive open-weight Qwen releases to turn model improvements into a broader distribution and buyer-choice strategy across sizes and deployment contexts.
Qwen3.8-27B can now be run locally! ✨ Run on 17GB RAM via Unsloth Dynamic GGUFs. Qwen3.8-27B is by far the strongest model for its size. We also uploaded NVFP4 quants. GGUF: https://huggingface.co/... Guide: https://unsloth.ai/...
@udiwertheimer
Udi Wertheimer on x
beats opus 4.6 and runs on a laptop 😂 dario you should've IPO'd last year 😂
@theahmadosman
Ahmad on x
Don't try to run Qwen 3.8 27B on a DGX Sparks / Mac minis / MacBooks Qwen 3.8 27B is a Dense model and those Unified Memory boxes want MoEs It might fit, but it will be very slow This model wants those 3090s, 5090s, RTX PRO 6000s, etc GPUs > Unified Memory for Dense models
@max_paperclips
Shannon Sands on x
fr though, how undertrained ARE models? How much improvement is still left on the table just from scaling data and quality? Do we eventually see Qwen 4-27b at Fable level? It's honestly kinda crazy. We're so far from Chinchilla now it's mad
@simonw
Simon Willison on x
The new Qwen 3.7 27B, running as a 17GB GGUF in LM Studio on my M5 Max laptop, just drew me the best pelican riding a bicycle I've seen from any model that runs on my laptop
@shuai_bai_
Shuai Bai on x
(Maybe) Opus 4.6-level agents, now local. Qwen3.8-27B is out. Try it and let us know how it goes 🎉
@sudoingx
@sudoingx on x
holy shit! look at the table qwen just published for the 27b. beating opus 4.6 max on computer use, 84.3 vs 72.7 on osworld. beating it on mobile use, 81.9 vs 62. beating it on multimodal software engineering. and visual math isn't even close, 94.6 vs 65.5. and i'll verify what
@teknium
@teknium on x
The new king of local models is out!
@maziyarpanahi
@maziyarpanahi on x
HOLLY 💩! 27B dense, native vision, 262K context... time to get it running fully local! images + hour-long video, Gated DeltaNet + gated attention, MTP, thinking control per request, and up to 1M context with YaRN. Qwen just made every RTX and DGX owner very happy today 😅
@itspaulai
Paul Couvert on x
I can't believe it Qwen3.8-27B is matching Opus 4.6 Max... the model that was the best (and the most expensive) just 6 months ago. And you can run it on your laptop. Locally. Fully open weights and under apache license. This level of intelligence in such a small model is sooo…
@gneubig
Graham Neubig on x
Third strong ~30B model coming out this week, amazing to see after having a bit of a drought.
@yagilb
@yagilb on x
Everyone is downloading this model right now. Send prayers to hugging face servers
@jumperz
@jumperz on x
Qwen 3.8-27B is finally here the jump from 3.6-27B is kinda insane... every single benchmark went up. Terminal coding: 63.4 → 73.0 SWE-bench Pro: 53.5 → 61.7 DeepSWE: 13.3 → 42.2 Software engineering: 49.3 → 79.0 remember you can run this on a $700 used 3090..
@aibattle_
@aibattle_ on x
Qwen 3.8-27B scores 42.2% on DeepSWE. Better than Gemini 3.5 Flash and close behind GLM-5.2 Max
@hesamation
@hesamation on x
you can basically run an Opus-4.6-ish model at 200 tok/s on a single RTX 5090.
@andrewcurran_
Andrew Curran on x
As promised, Qwen has released the weights for Qwen3.8-27B and they are up on Hugging Face.