Qwen3.8-27B (FR*EE) + Claude Code : The NEW King of Local AI

Tech / Product Summary (from video subtitles)

1) Model release & performance comparison

• Qwen 3.8 (27B) is presented as free-to-use when run inside Claude Code, via an endpoint.
• The speaker claims Qwen 3.8 (27B) outperforms Qwen 3.6 (27B) by a large margin across benchmarks:
  • TerminalBench v2.1: Qwen 3.8 (27B) 73 vs Qwen 3.6 (27B) (lower score mentioned; details not fully listed)
  • SWE-bench Pro: Qwen 3.8 (27B) 61.7 vs Sonnet 4.6: 58
• Additional side-by-side comparison vs Anthropic Sonnet 4.6 (with “medium” reasoning effort):
  • TerminalBench v2.1: Qwen 3.8 (27B) 73 vs Sonnet 4.6 67
  • SWE-bench Pro: Qwen 3.8 (27B) 61.7 vs Sonnet 4.6 58
• “Golden setup” recommendation for complex work:
  • Sonnet 4.6 + medium reasoning effort
  • They still prefer it over “Sonnet 5” because Sonnet 4.6 uses fewer tokens (lower cost) while still getting the job done.

2) Claimed core capabilities (from Hugging Face)

From the Hugging Face “features” highlighted in the video:

• Coding and “professional work”
• Research
• Long-horizon agentic tasks
• Stronger autonomous planning
• Better handling of environment feedback to improve end-to-end reliability
• Native multimodal capability (text + images + video)
• 27B dense architecture, described as smaller than a larger “3.8 max” model—making it more practical for:
  • Local deployment
  • Commercial/developer use
• Apache 2.0 license emphasized (open source / permissive)
• Suggested local requirement: run locally with ~24GB+ VRAM

3) How to use it in Claude Code (step-by-step via jan.ai)

The tutorial walks through wiring Qwen 3.8 (27B) into Claude Code using jan.ai:

1. Install the jan.ai desktop app (Windows / Mac / Linux)
2. In jan.ai, open Settings → Claude Code
3. Add a new provider:
   • Choose “API openAI compatible”
   • Provide a base URL from a Hugging Face-hosted free endpoint
   • Set key value to “none” (because the endpoint is claimed to require no key)
4. Refresh and select the Qwen 3.8 27B endpoint model
5. Map routing in Claude Code to Opus / Sonnet / Haiku style (speaker maps to Sonnet)
6. Example interaction:
   • Use a “/model” command to select 3.8 27B
   • Then run prompts and coding/debugging tests

4) Testing prompts and example workflows

The speaker describes testing prompt ideas generated via ChatGPT, including:

• Deep reasoning prompts
• Coding challenge prompts
• Debugging test prompts

They emphasize experimenting interactively to evaluate the model’s “taste” in agentic / coding tasks.

5) Vision / multimodal demo (image understanding & extraction)

A concrete vision test is performed:

• A mobile app prototype / Figma-like image (WebP) is uploaded from a local folder.
• The model is prompted (as a “senior engineer”) to:
  • Describe what it sees
  • Extract important numbers/levels
  • Separate observations vs assumptions
  • Not invent unseen information

The model is claimed to correctly read and extract:

• Time (9:41) (verified by opening the image in the browser)
• Battery level (96%)
• Vehicle information (e.g., Toyota, CHR 2026, as shown in the subtitles)
• Anomalies, plus recommendations/fixes

The interface indicates multiple detected screens:

• Screen 1: vehicle dashboard
• Screen 2: station map
• Screen 3: station details / vehicle details

6) Context window claims

• The model is described as supporting a 262K native context window
• The speaker claims it can be extended up to 1 million context.

---

Main speakers / sources

• Speaker: the video narrator (“Hey guys…”) presenting Qwen 3.8 (27B) and the Claude Code integration tutorial.
• Sources referenced:
  • Hugging Face (features page, free endpoint, model link)
  • jan.ai (desktop app used to connect to Claude Code)
  • Anthropic models: Claude Sonnet 4.6 (used as a comparison baseline)
  • “Victor” (identified as Head of Product at Hugging Face, credited with releasing the free endpoint)