SYLIN

Qwen

by Alibaba · Model · Language & code models

promising

Open Qwen 3.8 runs well on modest hardware, but it is wordy and independent testing is thin

Unsloth expects a smaller Qwen4-27B to run in 16 GB of RAM; users ask which quantization and how much context fits

AccessOpen weightsNeeds~17 GB; smallest build ~16 GBDownloadabout 17 GB for 3.8 weights; 14.3 GB for 3.6LicenseApache 2.0

Why we say so

Artificial Analysis rates Qwen 3.8 27B as more intelligent than Nemotron 3.5 Super, though slower. HN users run it locally on 16 to 32 GB: one replaced about 95% of their conversational Claude use, another said it is verbose, uses about twice the tokens of GLM-5.2, and still keeps DeepSeek 4.1 Flash for real work. Rivals' claims of beating Qwen 3.6 or 3.8 come from the makers themselves; Fahd Mirza found the unreleased Qwen 4 Max better at one image task than 3.8.

Promised Nov 16: open-source release 'very soon'

Versions

VersionReleasedRank
Qwen 3.8LATEST
promising Capable and runs on a 16 to 32 GB GPU, but verbose
Aug 14
Qwen 3.6
unproven Mostly appears as the baseline rivals claim to beat
Qwen 3.5
unproven Used as a base for other models; small-model research and speed notes only

Evidence

  1. Nov 19
    Weight edits lifted the 0.8B model's MATH-500 from 7.9% to 48%, mostly by fixing repetitive reasoning loops
  2. Nov 19
    Local Qwen3.8-27B Q4 covers leisure coding, but one user still uses DeepSeek 4.1 Flash for real work
  3. Nov 16
    Qwen says smaller Qwen4 models are coming to open source very soon; reviewer expects a 27B
  4. Nov 14
    Qwen3.8-27B UD-Q2_K_XL v3.1 roughly matches v3.0 on a home benchmark: slightly worse at debugging, 46.5k vs 61.7k tokens per task, fits 16 GB GPU
  5. Nov 13
    Blog post projects final-layer states of Qwen3.5-2B on 3,690 conversations; Chinese clusters apart from English, German, Croatian and Python overlap English
Show all 27
  1. Nov 11
    In ComfyUI, Qwen3.5 9B rose from 106 to 219 tok/s on RTX 5090 and 91 to 129 tok/s on RTX 5080 (v0.40.0 to v0.45.0)
  2. Nov 11
    Qwen3.8 27B describing a 1536x1536 image went from 55 to 122 tok/s on RTX 5090 in ComfyUI
  3. Sep 27
    Qwen3.8-27B: 64 layers, gated delta attention with full attention every fourth block, 262K context extendable to 1M, native image and video input, Apache 2.0, thinking on by default
  4. Sep 9
    Prefilled with GPT-5.5 Pro reasoning openings, Qwen 3.8 answers moved +20.58 points toward GPT-5.5 Pro, suggesting training on GPT traces
  5. Sep 9
    In an earlier experiment Qwen barely moved toward Opus 4.8 when prefilled with its reasoning, unlike its shift toward GPT-5.5 Pro
  6. Sep 7
    Agent 'Quinn' sent 50 unsolicited invoices totaling $12,350 and 113 spam emails via Mailjet; run halted early
  7. Sep 6
    Same coding task: Qwen3.8-27B on Cerebras burned $1.10 in 90 s unfinished; DeepSeek-V4-Flash finished in 172 s for $0.024
  8. Sep 3
    Qwen 3.8 27B now served on Cerebras at about 1500 tokens/s, billed per token
  9. Aug 28
    Q4_K_M (17GB) generates ~14 tok/s on Mac Studio M3 Ultra, about half of Qwen 3.6 27B's speed in the same setup
  10. Aug 26
    A 4-bit quant of Qwen3.8 27B reportedly matched Gemini 3.7 Flash in internal tests; ~140 tok/s single stream, ~800 tok/s at c=8 on RTX 5090
  11. Aug 25
    Flash-Next previews the architecture of the upcoming Qwen4 family; commenters hope the full family includes small models like 4B
  12. Aug 23
    In one commenter's reverse-engineering benchmark, DeepSeek V4 Flash did much better than Qwen 3.8 27B.
  13. Aug 16
    Qwen3.8 27B dense multimodal model, up to 1M context; claimed to beat Opus 4.6 Max on agentic coding and knowledge benchmarks
  14. Aug 15
    A local Qwen 27B model via Ollama in DSH built a working animated duck HTML page, but over-thinks, he says
  15. Aug 14
    Needed far less hand-holding than DeepSeek V4 Flash and Qwen 3.6 27B on the same agentic task
  16. Aug 14
    Claimed SWE-bench Pro rises from 53 to nearly 62 and OSWorld from 64 to 84 over the previous generation
  17. Aug 14
    Single-file HTML grilling-world page was decent for 27B size, but flames looked weaker than Opus 5 output
  18. Aug 10
    Qwen 27B still beats Muse Glimmer on computer use, terminal work and SWE-bench in Meta's table
  19. Aug 3
    Qwen 3.6 27B served through Ollama drove Hermes Agent note-taking tasks end to end on one GPU
  20. Aug 3
    Qwen3.8-27B dense model announced with open weights on Hugging Face under Apache 2.0
  21. Jul 20
    Alibaba released Qwen3.8 within days of Kimi K3, per the reviewer
  22. Jul 20
    Fixed the same bug with 45 API calls, 42 tool calls, fewer tokens and under half the time; stayed scoped and stopped cleanly

Cost to run

Alternatives

GemmaCommenter asks whether Gemma 4 E4B would be more practical than 3B-active Qwen MoE on a Pi
OrnithA 64GB M4 Max user switched to the MLX 4-bit build as daily driver, finding it better than Qwen3.6 35B-A3B
KimiBug-fix agent task: K3 found and fixed the planted bug plus three extra real bugs, but over-scoped and looped on a broken judge
Qwen Flash NextRuns about twice as fast as Qwen 3.8 27B on a quad-3090 rig; reviewer says it may replace the 27B
Muse GlimmerBeats Qwen 27B and a second rival on agentic work (MCP orchestration, deep search, banking, long-context recall)