Qwen3.7-Max: The Agent Frontier

Original Article ↗ Hacker News Discussion ↗

Benchmarks, Comparisons, and Marketing

  • Multiple commenters criticize Qwen for benchmarking against older competitors (e.g., Opus 4.6) instead of the latest GPT/Claude/Gemini versions, viewing it as marketing/expectation management.
  • Some note mismatches between benchmark results and lived experience: Qwen often looks stronger on paper than in day‑to‑day use.
  • Others point out benchmarking lag (e.g., new models not yet in public eval suites).

Model Quality vs Frontier Models

  • Many see Qwen as “great for open weights,” close to frontier but not equal to top proprietary models.
  • Some users report Qwen 3.6 (especially 27B) being good enough to replace mid‑tier models (e.g., Sonnet‑class) for many coding tasks, but still below top‑end (Opus/GPT) on complex work.
  • There is anecdotal debate about Anthropic model regressions (4.6 vs 4.7), with conflicting reports and claims of “nerfing” vs harness issues.

Open vs Proprietary & Hosting

  • Qwen3.7‑Max is proprietary; “Plus/Max” lines are generally not open weights. People hope for later open releases or 3.7 analogs to 3.6 open models.
  • Some want Qwen hosted by US‑based providers (e.g., Fireworks/OpenRouter/Friendli) for compliance and latency; others note 3.6 Plus already appears via some US proxies.

Local Deployment, Hardware, and Performance

  • Extensive discussion on running Qwen 3.6 locally:
    • 27B dense vs 35B A3B MoE trade‑offs: dense is smarter but slower; MoE is much faster with slightly lower quality.
    • Quantization choices (Q4/Q5/Q6, K_M vs K_XL) and KV‑cache strategies strongly affect speed and quality.
    • New MTP (multi‑token prediction) variants can roughly double generation speed in some setups.
  • Hardware advice spans M‑series Macs (esp. 64–128GB), Strix Halo boxes, RTX 6000, multi‑GPU rigs, and budget cards (3060, P40), with cost ranges from ~$2.5k laptops to $10k+ GPU builds.
  • Consensus: local models are slower and more work to tune but provide privacy and predictable costs.

Coding Agents and Tooling

  • Several users report good results using Qwen 3.6 (27B or 35B) with coding agents such as pi, Claude Code as a harness, OpenCode, and VS Code integrations.
  • Proper harness configuration (large context, “preserve_thinking”, tools) significantly impacts effectiveness.
  • For many, Qwen 3.6 is “good enough” to offload a substantial share of coding tasks from paid frontier models.

Hallucinations, Token Efficiency, and Metrics

  • Qwen3.7‑Max is highlighted as SOTA on Artificial Analysis’s “non‑hallucination rate” for omniscience, but commenters stress this metric alone is insufficient:
    • A model can avoid hallucinations by refusing to answer.
    • The Omniscience Index (which balances correctness, refusals, and hallucinations) is viewed as more meaningful.
  • Token efficiency becomes a major concern:
    • Some Chinese and Nvidia‑branded models are criticized for needing many more tokens to reach similar performance.
    • Gemma 4 is cited as an example of high token efficiency; Qwen and DeepSeek are sometimes described as “chatty.”
    • Users want models that stay close to frontier quality but minimize tokens for cost and latency.

Censorship, Geopolitics, and Trust

  • Several commenters won’t use Chinese‑hosted models for corporate or sensitive work, fearing government access and IP exfiltration.
  • Others argue US services pose similar surveillance risks; the debate extends to PRISM‑era programs and global intelligence cooperation.
  • Concrete examples:
    • Hosted Qwen models reportedly refuse to discuss Tiananmen or Uyghurs, while “decensored” local variants do.
    • Western models are said to provide more detailed answers on some controversial topics but still show alignment/censorship on others.
  • Some Europeans feel stuck between distrusting both US and Chinese providers and lacking strong native alternatives.

Economics, Cloud vs Local, and ROI

  • Opinions split on whether to rent GPUs (Runpod/Vast) vs buy hardware:
    • Rentals are often priced to pay off hardware in 1–1.5 years; some argue buying is better if usage is sustained.
    • Others note that “other people’s compute” is simpler and avoids large capex but sacrifices privacy and control.
  • Individual anecdotes:
    • Heavy users saw $100–200/month SaaS AI bills, which nudged them toward buying high‑RAM laptops to run local Qwen/Gemma.
    • Some compare local‑LLM expenditure to the 2016–2018 crypto GPU wave, warning about hype vs real productivity.

Adoption, Alternatives, and Switching Behavior

  • Noticeable migration patterns:
    • Some users report moving from Google Pro (Gemini/Flash) to Qwen/DeepSeek due to pricing and quota limits.
    • Others shifted from Claude (especially dissatisfaction with 4.7) to Kimi, Qwen, or DeepSeek models.
  • People see a rough split:
    • Frontier APIs for highest‑stakes or hardest problems.
    • Open or cheaper models (like Qwen 3.x) for everyday coding and planning, often via local or budget hosting.