Qwen 3.8 27B Review — A 27B Dense Model That Beats Opus 4.6 Max on Agentic Coding Benchmarks
✅ Pros
- • DeepSWE 1.1 score of 42.2 beats Opus 4.7 Max's 40.0 — a 27B open model outperforming Anthropic's frontier on the same Claude Code harness, confirmed by HN users
- • Runs on consumer hardware: 27B dense with FP8/NVFP4/GGUF quants — HN users running it on RTX 4090, RTX 6000, and Strix Halo APUs within hours of release
- • True vision-language model (image-text-to-text) with video understanding — OSWorld-Verified 84.3, WebArena-Verified 64.8, AndroidWorld 81.9 beat much larger models
- • 262K native context extensible to 1M; flexible thinking control (reasoning_effort, preserve_thinking) — thinking on by default, disable per request
- • Apache-2.0 license, FP8 official weights, and instant ecosystem support: Unsloth GGUF/NVFP4 quants plus DGX Spark vLLM configs published the same day
- • Big generational jump: QwenSWEBench 79.0 vs 3.6's 49.3, CoWorkBench 70.7 vs 61.0, IFBench 79.5 vs 69.1
⚠️ Cons
- • Factual recall is compressed: HN users note 27B-class models are worse than GPT-3.5-era models at recalling facts — 'you can't compress the entire human knowledge into a 30GB file' (redox99)
- • Benchmark comparisons use the Claude Code harness; HN skeptics argue open-model benchmarks overstate real-world gains ('They do not beat opus on real-world usage' — NitpickLawyer)
- • Quantized models can lose focus after long context — 'quantized model lose focus after long context and can do damages or thinking loop' (ycui7)
- • Qwen Cloud hosted version is 'coming soon' — no official managed API yet; the FP8 release is the only official artifact right now
- • At 27B with 262K context, memory requirements are steep for laptops: full FP8 needs ~30GB VRAM; useful quants need 16-24GB
Developers and teams who want frontier-adjacent agentic coding and multimodal capability on consumer/on-prem hardware at open-weights cost — 'the best compromise between size and intelligence to run on consumer hardware'
Free (Apache-2.0 open weights, FP8 on Hugging Face); Qwen Cloud hosted API coming soon; local inference cost = hardware only
Quick Verdict
Qwen3.8-27B is the release that made HN’s local-model crowd sit up: a 27B dense vision-language model whose agentic coding scores beat models 10x its size. On DeepSWE 1.1 it scores 42.2 — ahead of Opus 4.7 Max’s 40.0 in Qwen’s own harness-controlled eval — and on Terminal Bench 2.1 (Terminus) it hits 73.0, within striking distance of Opus 4.6 Max’s 78.2. The multimodal scores are the sleeper: OSWorld-Verified 84.3 beats every model in Qwen’s comparison table including Opus 4.6 Max (72.7).
For a model that runs on a single RTX 4090-class GPU in quantized form, that is remarkable. It is also the most complete small-model release of the week: official FP8 weights on Hugging Face (Apache-2.0), Unsloth GGUF and NVFP4 quants up the same day, and DGX Spark serving configs published within hours. The 794-point HN thread largely reads as a buying guide for local inference hardware.
Rating: 8.1/10 — the best 27B-class open model yet, with the caveat that benchmark victories over Opus don’t fully survive contact with real-world workflows.
Architecture & Release
Qwen3.8 is the newest generation of Qwen’s open family, built on the Qwen3.5 architectural foundation. The 27B is a dense model with a vision encoder: a native vision-language model that understands images and videos, from STEM diagrams to hour-scale video. Key specs:
- Parameters: 27B dense (hidden 5120, 64 layers)
- Hybrid attention: Gated DeltaNet (48 V heads / 16 QK heads) interleaved with Gated Attention (24 Q heads / 4 KV heads) — 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
- Context: 262,144 native, extensible to 1,000,000
- Thinking control: thinking on by default, disable per request;
reasoning_efforttunes depth;preserve_thinkingretains reasoning context across turns - MTP: multi-token prediction trained
- Quantization: official fine-grained FP8 (block size 128) with “performance metrics nearly identical to the original model”
- License: Apache-2.0
The FP8 release (Qwen3.8-27B-FP8) is compatible with Transformers, vLLM, SGLang, and TokenSpeed. Unsloth’s GGUF quants run on “pretty much every single potato” (benxh), while the NVFP4 variant targets Blackwell’s native W4A4 compute path — one user reports 1.2-1.5x performance gains over FP8 on RTX 6000 for high-concurrency enrichment workloads.
Benchmarks: The Numbers That Matter
| Benchmark | Qwen3.8-27B | Qwen3.6-27B | Qwen3.7-Plus | Muse Glimmer-30B | Opus4.6 Max |
|---|---|---|---|---|---|
| Terminal Bench 2.1 (Terminus) | 73.0 | 63.4 | 64.0 | 51.7 | 78.2 |
| SWE-bench Pro | 61.7 | 53.5 | 57.6 | 51.2 | 53.4 |
| DeepSWE 1.1 | 42.2 | 13.3 | 14.2 | — | — |
| QwenSWEBench | 79.0 | 49.3 | 59.2 | — | 63.8 |
| LiveCodeBench v6 | 90.3 | 83.9 | 89.6 | — | 88.8 |
| GPQA Diamond | 89.2 | 87.8 | 90.3 | 83.5 | 91.3 |
| IFBench | 79.5 | 69.1 | 79.1 | 77.0 | 62.5 |
| OSWorld-Verified | 84.3 | 63.9 | 73.3 | 65.9 | 72.7 |
| WebArena-Verified | 64.8 | 48.8 | 55.3 | — | — |
| AndroidWorld | 81.9 | 70.3 | 81.0 | — | 62.0 |
| OmniDocBench 1.5 | 91.1 | 89.4 | 91.4 | 75.8 | 86.6 |
Caveats Qwen itself publishes: SWE-bench Pro and DeepSWE use the Claude Code harness at temp=1.0, top_p=0.95, 256K context; NL2Repo-Bench disables repository-accessing bash commands to prevent reward hacking; DeepSWE and SWE-bench results are harness-dependent, which HN skeptics were quick to point out.
The most striking cell: DeepSWE 1.1 42.2 vs Opus 4.7 Max 40.0 (per HN user scrlk: “Beats Opus 4.7 Max (w/ Claude Code) on DeepSWE (42.2 vs 40). Looks like Qwen’s 27B models continue to pack some punch.”). And on OSWorld-Verified the 27B beats every comparison model including models with dedicated computer-use training.
Deployment Reality
The HN thread confirms the model runs where the community actually has hardware:
- DGX Spark: vLLM NVFP4 configs published by Erdal Toprak the same day
- RTX 4090 / RTX 6000: Unsloth GGUF and NVFP4 quants — “This is the version we’ll be testing on our rtx 6000 today!” (hadlock)
- Strix Halo (AMD APU): “My Strix Halo is about to go overdrive!” (brcmthrowaway) — the 27B density is sized for unified-memory laptops
- llama.cpp: GGUF quants from Unsloth for any CPU/GPU combo
Practical guidance from the thread: official FP8 safetensors for vLLM, Unsloth GGUF for llama.cpp, NVFP4 for Blackwell high-concurrency workloads. One user warns: “if you have the VRAM, use official release. quantized model lose focus after long context and can do damages or thinking loop” (ycui7).
Community Verdict
The thread (794 points, 520 comments) was unusually practical — serving configs, quant comparisons, and hardware talk. The optimistic read: “These are massive improvements - and something you can actually run on a laptop” (chvid); “27b dense model at Opus 4.6 level… Opus at home” (tosh). The counter-read from NitpickLawyer: “They do not beat opus on real-world usage… Qwen models are impressively good for what they are, are ‘good enough’ for plenty tasks, can be ran locally on decently priced hardware, and so on.” A telling data point from an embedded-systems shop: Qwen3.6-27B scored only 4% below Opus-4.8 pass@1 on their Rust/C mmWave radio repos (n=250, paired results).
The factual-recall critique is worth noting: “These small models are actually worse than GPT 3.5 at some tasks (like recalling facts)… you can’t compress the entire human knowledge into a 30GB file” (redox99) — with the retort that web search is the right tool for facts, and the model’s job is reasoning. Quantization transparency also drew fire: KL divergence is “a barely useful signal” (zargon) compared to real benchmarks.
Alternatives
- Qwen3.6-27B — The previous generation; still solid, and a good budget pick, but the 3.8 generation is +10-30 points across coding and agent benchmarks.
- Muse Glimmer-30B — The main rival 30B open model; Qwen3.8-27B beats it on every shared benchmark (IFBench 79.5 vs 77.0, OSWorld 84.3 vs 65.9).
- Opus 4.6/4.7 Max (Anthropic) — Still ahead on Terminal Bench (78.2) and GPQA (91.3), and better on real-world nuance per skeptics; costs API rates and carries no open weights.
- DeepSeek V4 Flash — Cheaper per token for text workloads; weaker vision and agentic coding; better for high-volume text pipelines.
- GLM-5.3 (Z.ai) — Frontier-class agentic coding at 5.2+ scale; far larger model, not runnable on consumer hardware, and text-only.
FAQ
Q: Can I run Qwen3.8-27B on my laptop? A: Yes, in quantized form. Unsloth GGUF quants run on llama.cpp across “pretty much every potato”; the NVFP4 variant targets Blackwell GPUs (RTX 6000-class) with 1.2-1.5x throughput gains; users confirmed Strix Halo APUs handle it. Full FP8 needs ~30GB VRAM.
Q: Is Qwen3.8-27B multimodal? A: Yes — it’s a native vision-language model (image-text-to-text) that understands images and video, including hour-scale video, STEM diagrams, and documents (OmniDocBench 1.5: 91.1).
Q: What’s the license? A: Apache-2.0. The official release is Qwen3.8-27B-FP8 on Hugging Face; the base BF16 model is Qwen/Qwen3.8-27B.
Q: Does it really beat Opus? A: On specific harness-controlled benchmarks, yes — DeepSWE 1.1 42.2 vs Opus 4.7 Max 40.0, OSWorld-Verified 84.3 vs Opus 4.6 Max 72.7. HN practitioners caution that real-world coding nuance still favors Opus; several report small gaps (4%) on their own evals.
Q: What context length does it support?
A: 262,144 tokens natively, extensible to 1,000,000. Thinking mode is on by default and can be disabled per request with reasoning_effort control.
Q: Is there a hosted API? A: Qwen Cloud will offer Qwen3.8-27B with 1M context and built-in tools — “coming soon.” Until then, it’s self-hosted (vLLM/SGLang/llama.cpp) or via community providers.