DeepSeek v4 Flash + Hermes Agent = Surprisingly STRONG

DeepSeek V4 Flash is back free on Nous Portal and ranks #10 of 87 models

The creator's push: stop chasing flagship models for every agent task. On Nous Portal, DeepSeek V4 Flash is the most-used model this month and the highest-consumed token on Hermes Agent — and it's currently free, even on the free subscription plan.

Benchmarks

V4 Flash sits at intelligence rank #10 out of 87 models, behind Gemini 2.6 (#1) and V4 Pro (#5). Speed rank is #4 at ~134 output tokens/second. Context window is 1M tokens with hybrid attention. The creator's framing: a lightweight model in the top 10 "speaks a lot of volume about its competence."

V4 Flash vs V4 Pro

A power user test: V4 Flash wins ~35% of real tasks while costing ~4x less than V4 Pro. On five coding problems, Pro spent 4x more output tokens but produced identical answers.

What it does well (the 5-test run)

• Python refactor (453-line data collector): no rate limit hit, identified hardcoded API keys in one pass.
• Repo mapping (FastAPI GitHub repo): completed in ~12 min at ~1,900 tokens. Local 50-file project finished in ~1 min at ~1,200 tokens.
• Einstein's zebra puzzle: passed with long logical chains, no fact-check by creator.
• Long-context summarization: extracted top 5 action items from 3 academic papers (2 PDF, 1 HTML) in ~2 minutes, output a merged MD with cross-cutting themes.
• Sub-agent delegation: 3 parallel research sub-agents spawned on its own (Python async, SQLite vs PostgreSQL, HTMX vs React) — output was solid but took 14 minutes.

Where it loses

Don't use V4 Flash for full code base analysis in one shot, vision/image tasks ("pretty inconsistent"), one-shot prompts, creative writing, or orchestration. Sub-agent speed is "way too long" — feed long docs directly instead of delegating. Verbosity is high, but the creator treats that as logical justification, not a flaw.