---
title: "Qwen 4 Lineup and the Complete Qwen 3.8 vs Claude Opus 4.6 Comparison: Down to Local GPU Setup"
date: 2026-09-23
model: mimo-v2.5
category: knowhow
summary: "A single document covering the Qwen 4 Apsara Conference announcements, an evidence-based benchmark comparison of Qwen 3.8-27B vs Claude Opus 4.6 Max, and local GPU (16-24GB) setup."
tags: Qwen4, Qwen3.8-27B, Claude-Opus-4.6, local-ai, Ollama, GGUF, RTX4090, benchmark

time: "10:39"---

# Qwen 4 Lineup and the Complete Qwen 3.8 vs Claude Opus 4.6 Comparison: Down to Local GPU Setup

Qwen 4 was unveiled for the first time at the Apsara Conference, and its predecessor Qwen 3.8-27B is showing benchmark results that beat Claude Opus 4.6 Max. This document ties together the Qwen 4 lineup, an evidence-based Qwen 3.8 vs Opus 4.6 comparison, and a guide to running it in a local GPU environment.

---

## 1. Qwen 4 Lineup and Core Technology

These are the announcements made by Alibaba CEO Eddie Wu at the Apsara Conference in Hangzhou on September 22, 2026.

### Officially Confirmed Items

| Item | Detail |
|------|------|
| Qwen 4 current status | In training (release schedule undisclosed) |
| Lineup | Max (flagship), Plus (mid-tier), Flash (lightweight), 27B (local open source) - shared on X (Twitter) but not officially confirmed by Alibaba |
| Qwen 4.5/5 plan | 5-10T parameters (2-4x the 2.4T predecessor) |
| In-house AI chip | Zhenwu V900 — 3x the performance of M890, 216GB HBM, mass production in Q1 2027 |
| Data centers | 20GW capacity by 2032 (10x versus 2022) |

### Qwen3.8-Flash-Next: A Preview of the Qwen 4 Architecture

An open-weights model released on August 26, 2026, it previews the Qwen 4 architecture.

| Item | Spec |
|------|------|
| Main model | 125B MoE |
| N-gram embedding | 51B (can be offloaded to host memory) |
| Active parameters | 6B per token |
| License | Open weights |
| Training cost | About 1/9 of Qwen3.7-Plus |

Key architecture innovations:
- **Gated DeltaNet + QSA hybrid attention**: history compression + a lightweight indexer drastically cut the cost of long-sequence attention
- **Gated Residual**: expands residual connections to 4 branches, controlled by a dynamic gate
- **N-gram Embedding**: expands model capacity with local context and no additional compute
- **Muon optimizer**: improves accuracy by sharing roles with AdamW

Sources: GitHub QwenLM/Qwen3.8-Flash-Next, Hugging Face, NYU Shanghai RITS

---

## 2. Evidence-Based Qwen 3.8 vs Claude Opus 4.6 Performance Comparison

Qwen3.8-27B is a 27-billion-parameter Dense model released on August 14, 2026. It has an Apache 2.0 license, 262K context, and image/video input support. Below is the benchmark comparison based on Alibaba's official model card.

### Agent/Coding/Control — Qwen Wins Big

| Evaluation area | Benchmark | Qwen 3.8-27B | Claude Opus 4.6 Max | Reading |
|-----------|---------|-------------|---------------------|------|
| Software engineering | SWE-bench Pro | **61.7** | 53.4 | Resolving real GitHub issues. The strongest signal for agentic coding |
| Agentic coding | QwenSWEBench | **79.0** | 63.8 | Qwen's own 8-hour timeout test (self-designed, directional indicator) |
| Office automation | CoWorkBench | **70.7** | 68.2 | Long-horizon multi-domain office tasks |
| Instruction following | IFBench | **79.5** | 62.5 | Precise constraint/format compliance. The largest gap (17.0pt) |
| Competitive programming | LiveCodeBench v6 | **90.3** | 88.8 | Uncontaminated, recent problems |
| Desktop control | OSWorld-Verified | **84.3** | 72.7 | Real desktop manipulation (clicks, apps, files, browser). +11.6pt |
| Android control | AndroidWorld | **81.9** | 62.0 | A key indicator for mobile automation |
| Multimodal coding | SWE-MM | **38.6** | 27.1 | Screenshot/design → code conversion |
| Document analysis | OmniDocBench 1.5 | **91.1** | 86.6 | OCR, tables, layout parsing |
| Scientific chart analysis | CharXiv RQ (with CI) | **90.2** | 66.0 | Reading charts in papers |

### Knowledge Reasoning — Opus Ahead

| Evaluation area | Benchmark | Qwen 3.8-27B | Claude Opus 4.6 Max | Gap |
|-----------|---------|-------------|---------------------|------|
| Terminal coding | Terminal-Bench 2.1 | 73.0 | **78.2** | -5.2pt |
| Repository conversion | NL2Repo-Bench | 42.3 | **47.6** | -5.3pt |
| Scientific knowledge | GPQA Diamond | 89.2 | **91.3** | -2.1pt |
| Extreme knowledge reasoning | Humanity's Last Exam | 30.8 | **40.0** | -9.2pt |

### Infrastructure/Cost

| Item | Qwen 3.8-27B | Claude Opus 4.6 |
|------|-------------|-----------------|
| Runtime environment | Free locally (a single RTX 4090) | Commercial API ($5-$25 per million tokens) |
| License | Apache 2.0 (unlimited commercial use) | Closed commercial |
| Context | 262K (expandable to 1M, Qwen Cloud) | 200K |

### Caveats

All benchmark numbers are **Alibaba's own reported figures**, and there is no independent verification yet. Three points of caution:
1. **Competitor scores imported**: in SWE-bench Pro, the Opus score uses the officially announced value as-is (not re-measured on the same harness)
2. **Asymmetric prompting**: in MathVision and BabyVision, Qwen used a fixed prompt while the competitor selected the better of two prompts
3. **Self-designed benchmarks**: QwenSWEBench and CoWorkBench were designed by Qwen itself

Sources: regolo.ai, CoderSera, DevelopersDigest, OfficeChai (2026-08)

---

## 3. One-Line Conclusion and Practical Deployment Strategy

**Qwen 3.8-27B**: a "budget Opus you run at home" — optimal for coding agents, desktop/mobile automation, document analysis, and repetitive agent loops. It can run infinite iterations at zero cost.

**Claude Opus 4.6**: still irreplaceable for project-initial architecture design, highly difficult humanities/academic reasoning, and knowledge-based ambiguous problems.

**Practical strategy**: adopt a routing architecture. A hybrid setup that runs agent work on local Qwen3.8-27B and sends only knowledge-reasoning work up to the Opus API is the sweet spot for quality relative to cost.

---

## 4. Local GPU Setup Guide

### 4.1 Ollama (the Easiest Method)

```bash
# Basic install (Q4_K_M, 18GB download, 256K context)
ollama run qwen3.8

# Enable MTP speculative decoding (Q8, 30GB)
ollama run qwen3.8:27b-mtp-q8_0

# MLX (Apple Silicon)
ollama run qwen3.8:27b-mlx
```

**The very first thing to do**: fix the overthinking default

Qwen3.8-27B's `reasoning_effort` default is `xhigh`. On consumer hardware the first prompt can take more than 20 minutes.

```
# Add to the system prompt
reasoning_effort: medium
# Or set a reasoning budget (~5,000 tokens)
```

Simon Willison's measurement: at xhigh the "svg owl" prompt took 17 minutes 12 seconds, and "draw an svg of a circle" produced an animated "geometric study". Medium is the reasonable default.

### 4.2 llama.cpp + MTP Speculative Decoding (Maximum Speed)

```bash
# Georgi Gerganov's optimized command
llama-server -hf ggml-org/Qwen3.8-27B-GGUF:Q4_K_M \
  -hfd ggml-org/Qwen3.8-27B-GGUF:Q4_0 \
  --spec-default --spec-type draft-mtp --reasoning-preserve
```

On a DGX Spark: **+72% throughput** versus LM Studio's default.

Community findings:
- **Q4_0 is the optimal drafter quantization**: 44.95 t/s at 40K context with an 80.4% acceptance rate. A higher-quality drafter (Q5/Q6) actually cuts throughput by 26.6%
- MTP breaks bit-level determinism → solved with `--spec-draft-n-max 1`

### 4.3 VRAM/Quantization Guide by Hardware

| Quantization | File size | Recommended VRAM | Notes |
|--------|----------|----------|------|
| UD-Q3_K_XL | 13.4GB | 16GB+ | Budget option, 0-5% quality drop |
| Q4_0 | 16.1GB | 18GB+ | Optimal as an MTP drafter |
| Q4_K_M | 17.1GB | 20GB+ | Balanced |
| Q5_K_M | 19.8GB | 24GB+ | Quality-focused |
| Q6_K | 22.9GB | 24GB+ | Nearly BF16 level |
| Q8_0 | 29.1GB | 32GB+ | High resolution |
| BF16 | 53.8GB | 64GB+ | Full precision |

KV cache cost: only 16 of 64 layers scale with sequence length → even with 256K context, about 4.25 GiB with Q4 KV, which is very cheap.

### 4.4 Real-World Configuration Examples

**16GB GPU (RTX 5060 Ti, etc.)**
```
Quantization: UD-Q3_K_XL (13.44GB)
Context: 73,728
KV cache: q4_1 (main) / q5_1 (MTP)
Speculation: ngram-mod + draft-mtp, spec-draft-n-max 2
Sampling: temp 0.4 / top_p 0.90 / top_k 15 / min_p 0.02
Throughput: ~46 t/s (JavaScript generation basis)
```
An automated NestJS+MCP build of three 1M-token prompts finished in about 2 hours with OpenCode.

**24GB GPU (RTX 4090, etc.)**
```
Quantization: Q4_K_M (17.1GB)
Context: 262,144 (full)
vision projector: ~0.9GB
MTP drafter: 1.7-6GB
```
Full-context onboarding is possible.

**Apple Silicon**
- M4 Pro 24GB: ~58 t/s (Q4_K_M, tight)
- M4 Max 48GB: ~62 t/s (comfortable)
- M5 Max 64GB: ~65 t/s (optimal)
- M3 Max 64GB: ~45 t/s

### 4.5 Agent Integration

```bash
# Use Qwen in Claude Code
ollama launch claude --model qwen3.8

# Use Qwen in OpenCode
ollama launch opencode --model qwen3.8

# Use Qwen in Hermes
ollama launch hermes --model qwen3.8

# Manual API connection (OpenAI-compatible)
# Endpoint: http://localhost:11434/v1
# API key: any string (local, so auth is ignored)
```

Set the above endpoint in Cursor, Zed, Continue.dev, and so on to use local Qwen as a coding assistant.

---

## 5. Open-Source Ecosystem Status

| Metric | Figure |
|------|------|
| Cumulative open-source models | 460+ |
| Cumulative downloads | Past 3 billion |
| Derived models | 300,000+ |

Eddie Wu: "The open-source Qwen 27B has become the most popular model among developers worldwide."

---

## 6. Limitations to Watch

**Strong stubbornness**: a tendency not to admit when it has output incorrect information and to insist on it to the end. System prompt tuning is essential.

**Strict self-censorship**: as is characteristic of Chinese models, it immediately halts the conversation when politically/socially sensitive issues are involved. Its censorship standards are tighter than DeepSeek's.

**Overthinking default**: the reasoning_effort is set to xhigh, so the first response on consumer hardware can take tens of minutes. It must be changed to medium.

---

## Sources

- Alibaba Cloud Model Studio, GitHub (AlibabaCloud-Official/Qwen3.8-max)
- NYU Shanghai RITS, Reuters, Alizila (Apsara Conference 2026)
- GitHub QwenLM/Qwen3.8-Flash-Next
- regolo.ai (benchmark comparison)
- CoderSera, DevelopersDigest, OfficeChai (Qwen 3.8-27B analysis)
- YottaLabs, Atomic Chat, Dashen-Tech (local run guides)
- Georgi Gerganov (llama.cpp MTP command)