Update — September 1, 2026: Celeris-1 Magnus is a hosted agent-tuned derivative built from Qwen 3.8-27B — hybrid diffusion, τ³-bench banking lead, same $0.20/$0.70 API pricing as base Celeris-1. Open weights for Magnus are not confirmed; this post still covers the open-weight base model itself.
Update — August 25, 2026: A ModelScope leak points to Qwen3.8-Flash-Next — a 125B-A6B MoE on teased Qwen4 architecture, expected Aug 26 on ModelScope/Hugging Face. That is the MoE sibling many commenters asked for here; it is sized for 128GB Macs and offload runtimes, not the smaller 35B-A3B shape RTX 5090 owners wanted.
Update — August 24, 2026: FreeToken's 753B-on-one-GPU result demonstrates the other side of local inference: a sparse MoE can exceed VRAM by hundreds of gigabytes when the runtime treats host RAM and PCIe as part of the machine. Our hardware guide corrects the “single GPU” shorthand with the tested 512GiB workstation configuration.
Update — August 20, 2026: Unsloth shipped a full Dynamic v3.0 GGUF release for Qwen3.8-27B, with a 1-bit quant that runs on 8GB of RAM and a hardware table covering every quant tier.
Qwen3.8-27B-FP8 hit #1 on Hacker News today with 893 points — and it closes a gap our own coverage flagged two days ago. When Qwen3.8-Max's open weights landed on August 12, we noted the smaller, locally-runnable Qwen3.8-27B companion Alibaba promised alongside it was still missing. It's not missing anymore.
Unlike the 2.4-trillion-parameter Qwen3.8-Max flagship — which needs a GB300 NVL72 rack to serve at scale — Qwen3.8-27B is a dense 27-billion-parameter model that runs, quantized, on a single consumer GPU. That's the detail driving the Hacker News thread: a genuinely local model whose own maker's benchmark numbers put it within reach of Claude Opus-class scores on agentic coding tasks.
TL;DR — what people are asking
| Question | Direct answer |
|---|---|
| What shipped? | Qwen/Qwen3.8-27B-FP8 on Hugging Face — dense, vision-language, 262,144-token native context (1M via YaRN) |
| Is it really #1 on HN? | Yes — 893 points, one of the largest AI-model threads of the week |
| Can I run it on a 4090? | Yes, quantized — commenters report success on single RTX 3090/4090s, Mac Studios, even a DGX Spark |
| How fast? | Roughly 10–160 tokens/sec, depending on quant level and whether MTP/speculative decoding is on |
| Does it beat Claude Opus? | On one cited benchmark (SWE-bench Pro), narrowly. On several others (Terminal-Bench, GPQA Diamond, HLE) it trails — see the table below |
| Is the comparison trustworthy? | Take it as a vendor-reported data point, not a settled result — commenters flagged harness/temperature mismatches and "benchmaxxing" risk |
| Dense or MoE? | Dense — every parameter active per token, which is why people are asking for an MoE variant |
| How do I try it? | Unsloth GGUF quants via llama.cpp, LM Studio, vLLM, or SGLang — see the setup section below |
What actually changed vs Qwen3.6-27B
Qwen3.8-27B is the direct successor to Qwen 3.6 27B, which our earlier hands-on coverage found to be the local coding sweet spot — better agent output than the sparser Qwen3.6-35B-A3B, if noticeably slower. According to Alibaba's own model card, Qwen3.8-27B moves meaningfully past that baseline on agentic and coding-specific evals:
| Benchmark | Qwen3.6-27B | Qwen3.8-27B | "Opus4.6 Max" (cited) |
|---|---|---|---|
| Terminal-Bench 2.1 (Terminus) | 63.4 | 73.0 | 78.2 |
| SWE-bench Pro | 53.5 | 61.7 | 53.4 |
| QwenSWEBench | 49.3 | 79.0 | 63.8 |
| GPQA Diamond | 87.8 | 89.2 | 91.3 |
| HLE (Humanity's Last Exam) | 24.0 | 30.8 | 40.0 |
All figures come from the Qwen team's own model card comparisons, not independent explainx.ai testing. Two things jump out. First, the jump from Qwen3.6-27B to Qwen3.8-27B is large and consistent across every category — this is not an incremental point release. Second, on SWE-bench Pro specifically, Qwen3.8-27B's cited 61.7 edges past the cited Opus4.6 Max figure of 53.4 — the one line in the table doing most of the "rivals Claude Opus" work in the Hacker News headline.
New model capabilities beyond raw scores: flexible reasoning_effort (xhigh/medium/low), preserve_thinking across multi-turn conversations, and multi-token prediction (MTP) for faster inference — the same MTP mechanism that made Qwen 3.6 27B's local setup meaningfully faster on Apple Silicon.
Does it actually rival Claude Opus? Read the table, not the headline
The "27B model beats Claude Opus" framing circulating on Hacker News overstates what the numbers show. Qwen3.8-27B beats the cited Opus4.6 Max figure on one of the five benchmarks in Alibaba's own comparison (SWE-bench Pro) and trails it on the other four — sometimes by a wide margin (HLE: 30.8 vs 40.0). That's a genuinely impressive result for a 27B dense model going up against a frontier closed model on any metric, but it's not a general capability win.
Several commenters in the Hacker News thread pushed back on the comparison methodology directly: different agent harnesses, different temperature settings, and different prompting setups between how Qwen benchmarked itself and how the cited Opus numbers were originally produced make apples-to-apples claims shaky. There's also a recurring "benchmaxxing" concern — the suspicion that a model was trained or tuned specifically to top particular leaderboard numbers rather than to improve general capability. explainx.ai's own benchmark contamination coverage covers why that skepticism is usually warranted for any vendor-published benchmark table, Qwen's included.
The honest read: Qwen3.8-27B is a genuinely strong dense model that closes much of the gap to frontier closed models on agentic coding tasks specifically — narrower context, no independent verification, and one favorable cherry-picked metric doing a lot of the "beats Opus" narrative work. Treat the comparison as a data point worth taking seriously, not a settled leaderboard result.
Can you actually run it locally? Hardware and realistic speeds
This is the question that made the thread take off. Qwen3.8-27B ships as Qwen/Qwen3.8-27B-FP8 — FP8-quantized weights straight from Alibaba — and Unsloth followed almost immediately with GGUF quantizations at multiple bit depths.
| Setup reported on Hacker News | Approx. speed | Notes |
|---|---|---|
| Single RTX 3090/4090, aggressive GGUF quant | ~10–40 tok/s | Lower quant levels trade quality for VRAM headroom |
| RTX 4090, higher quant + MTP enabled | ~60–100 tok/s | MTP is the multi-token prediction feature carried over from Qwen3.6 |
| Mac Studio (M4 Max) | ~30–80 tok/s | Consistent with the Apple Silicon numbers our Qwen3.6 27B coverage measured |
| NVIDIA DGX Spark | up to ~160 tok/s | Highest reported figure in the thread, with speculative decoding tuned in |
The wide 10–160 tok/s range across similar-class hardware comes down almost entirely to quantization level and whether MTP or speculative decoding is actually configured — not just raw GPU horsepower. That matches the pattern our Qwen 3.6 27B local dev guide documented in detail: MTP is the single biggest lever for local throughput on this model family.
Setup: llama.cpp, LM Studio, vLLM
# Pattern: serve an Unsloth GGUF quant with llama.cpp (adjust quant level to your VRAM)
llama-server -hf unsloth/Qwen3.8-27B-GGUF:Q8_0 \
--spec-type draft-mtp \
-ngl 999 \
-fa on \
-c 65536 \
--port 8080
That exposes an OpenAI-compatible endpoint at http://127.0.0.1:8080/v1 you can point OpenCode, LM Studio, or any agent harness at. If you'd rather not touch a terminal, LM Studio's model browser indexes Unsloth's GGUF builds directly — search Qwen3.8-27B and pick a quant that fits your available RAM/VRAM. For server-side deployment, vllm serve or SGLang both support the FP8 checkpoint directly for teams with a full GPU rather than a quantized local build.
Dense vs MoE: why people are asking for a smaller variant anyway
The single most repeated request in the Hacker News thread was for a mixture-of-experts (MoE) variant like Qwen3.6-35B-A3B, which activates only a small fraction of its parameters per token. A dense model like Qwen3.8-27B activates all 27B parameters on every single token, which is exactly why it's slower per token than an MoE of similar total size — even though the earlier Qwen3.6-27B dense model actually produced better agent output than the faster Qwen3.6-35B-A3B MoE in prior hands-on testing.
That's the real tradeoff underneath the request: dense models tend to punch above their parameter count on quality, MoE models punch above their parameter count on speed and lower-VRAM accessibility. Qwen3.8-27B is Alibaba betting on quality-per-parameter for this release; a Qwen3.8 MoE, if it ships, would be the speed-optimized sibling for people on tighter hardware. Nothing announced yet confirms one is coming.
Overthinking: the "xhigh" mode complaint
Not every reaction was positive. A recurring complaint in the thread concerned Qwen3.8-27B's reasoning_effort: xhigh mode — several users reported 20 to 90 minute completions for simple prompts, including one describing generating a basic SVG. That's the tradeoff of exposing flexible reasoning effort as a user-facing dial: xhigh clearly buys deeper reasoning on hard problems, but it's easy to leave it set for tasks that don't need it, and the model doesn't appear to self-limit. Practical takeaway: default to medium or low reasoning effort for routine tasks, and reserve xhigh for problems you've already confirmed need it.
Update — August 17, 2026: Simon Willison's numbers put a size on the problem
Simon Willison's own hands-on review (437+ points on Hacker News) puts hard numbers behind the xhigh complaint. Running the model at its documented default — xhigh reasoning effort — on both a 128GB M5 Max MacBook Pro and an Nvidia DGX Spark, his classic "pelican riding a bicycle" SVG test took 21 minutes, burning 22,276 reasoning tokens to produce just 3,223 tokens of actual output — over 6 reasoning tokens for every 1 output token. Turning reasoning off for the same prompt dropped that to 137 seconds, a roughly 9x speedup, at the cost of a visibly worse result (a badly-shaped bicycle frame, no hands on the handlebars). The most striking example: prompted with nothing more than "draw an svg of a circle," xhigh spent several minutes producing paragraphs of reasoning about "geometric studies," Bauhaus aesthetics, and ambient motion before generating an elaborate animated multi-ring composition — technically beautiful, and not remotely what was asked for.
Willison's practical recommendation: ignore the xhigh default entirely and start new sessions on low or no-reasoning, escalating only for problems that demonstrably need it — the same "default to low/medium" guidance in the table above, now with concrete before/after numbers behind it. He also confirmed the model handles two things worth knowing for builders: it's genuinely strong at vision tasks like bounding-box detection (correctly localizing pelicans in a photo on a 0-1000 coordinate scale, then one-shotting a full HTML visualization tool for that JSON output — albeit an over-engineered one under xhigh), and it can drive a real coding agent loop — he ran it through the Pi harness against a live codebase and got a coherent, accurate answer to "how does auth work?" alongside working file edits.
On raw speed, Willison measured 15-30 tokens/second in LM Studio on both machines — slow next to hosted API models (Artificial Analysis clocks GPT-5.6 Sol at 74 tok/s and Luna at 184 tok/s). The most promising fix so far is Multi-Token Prediction (MTP), a technique baked into the model itself where a cheap draft mechanism guesses several tokens ahead and the main model verifies them in bulk. Running llama.cpp with --spec-type draft-mtp produced a ~72% throughput improvement over the default LM Studio GGUF in Willison's own benchmark — a meaningful chunk of the local-vs-hosted speed gap closed by inference-side tooling alone, not a smaller or dumber model.
The "gap to frontier" narrative, again
Alibaba is explicitly positioning Qwen3.8-27B as a free, locally-runnable alternative that narrows the lag to frontier closed models to roughly six months — reigniting the China open weights vs. US closed frontier conversation that's been running through most of this year's Qwen, DeepSeek, and Kimi releases. It's worth reading that debate alongside this launch rather than in isolation: the pattern across DeepSeek Flash and Alibaba's own Qwen3.8-Max open-weight release three days earlier is a genuine acceleration in open-weight capability, paired with real caveats each time — licensing terms, stripped features, or benchmark methodology questions like the ones raised here.
A hosted Qwen Cloud API option is also planned, per the Hacker News discussion, but was not fully live at the time of this post — "coming soon" is the extent of what's confirmed.
Builder checklist
| Do | Don't |
|---|---|
| Pick your quant level based on available VRAM, and confirm MTP is actually enabled before judging speed | Compare raw tok/s numbers across setups without checking quant level and MTP status first |
| Treat the Opus comparison as one vendor-reported benchmark, not a verdict | Repeat "beats Claude Opus" without the caveat that it's true on one of five cited metrics |
Default reasoning_effort to medium/low for routine tasks | Leave xhigh on by default and wonder why a simple prompt takes an hour |
| Watch for an official Qwen3.8 MoE variant if you need lower-VRAM speed over quality | Assume a dense 27B model will match an MoE's tokens/sec on the same card |
| Pull Unsloth's GGUF quants for local testing today | Wait for the hosted Qwen Cloud API if you want to try it now — it isn't fully live yet |
Bottom line
Qwen3.8-27B is real, it's live, and it earned its #1 spot on Hacker News — a dense, locally-runnable, vision-language model that meaningfully improves on Qwen3.6-27B and gets genuinely close to Claude Opus-class performance on agentic coding benchmarks, at least by Alibaba's own accounting. Run it locally today via Unsloth's GGUF quants if you have a decent consumer GPU or Apple Silicon Mac. Just hold the "beats Opus" claim loosely — it's true on one benchmark out of five cited, the comparison methodology is contested, and the overthinking problem in xhigh mode is a real, reported cost, not a rumor.
Related on explainx.ai
- Perplexity portable Computer on DGX Spark — Qwen 3.8 27B local option (Aug 26)
- Ornith-1.5's self-improvement claims vs. this model — an HN commenter's own benchmark puts Ornith-1.5-35B behind Qwen3.8-27B on Terminal-Bench 2.1, SWE-bench Pro, and DeepSWE
- Unsloth's Dynamic v3.0 GGUFs for Qwen3.8-27B — hardware table and quant guide — the full quantization release, including a 1-bit build that runs on 8GB RAM
- OrcaRouter ships an uncensored Qwen3.8-27B MLX build for Mac — a third-party abliterated derivative of this same model, for Apple Silicon
- Qwen3.8-Max open weights are live — stripped, relicensed, half-delivered — the 2.4T flagship half of this same August 3 promise
- Qwen 3.6 27B local dev guide: llama.cpp, OpenCode, dense vs MoE — the predecessor this model directly improves on
- Unsloth Desktop: one local app that trains and runs models
- Top 10 open-weight models for a laptop
- Choosing open weight vs. closed AI models
- "American AI is losing" — the open-weights op-ed that split Hacker News
- Goodhart's Law and AI benchmark contamination
- DeepSeek Flash — 8 trillion tokens/day on OpenCode
Primary sources: Qwen/Qwen3.8-27B-FP8 on Hugging Face · Hacker News discussion thread (893 points, August 15, 2026)
Status as of August 15, 2026. Benchmark figures and hardware speed reports are as cited by the Qwen team's model card and the Hacker News community discussion at publication time — independent verification of the Opus comparison methodology was not available as of this writing. Follow @explainx_ai for updates, including if a Qwen3.8 MoE variant or the hosted Qwen Cloud API ships.