Do I need to update my Hermes agent to get the DeepSeek V4 Flash 0731 build?

No. The 0731 refresh kept the model ID deepseek-v4-flash, so any agent already calling it is served the latest build with no redeploy or config change.

Should a Hermes agent use DeepSeek V4 Flash or V4 Pro?

Flash for most work. It is about three times cheaper on output, carries five times the concurrency, and after the 0731 retraining matches or beats V4-Pro-Preview on several published benchmarks. Use Pro only for the hardest long-horizon reasoning.

What context window does DeepSeek V4 Flash have?

A 1 million token context window with a maximum output length of 384K tokens.

← Home

Guide

Hermes Agent + DeepSeek V4 Flash: The 0731 Build

DeepSeek refreshed V4 Flash on 31 July 2026. Same architecture, same parameter count, rebuilt post-training — and a jump on agent benchmarks large enough to change which model a cost-conscious Hermes Agent should default to. Here is how to run it, and the routing detail that decides whether it works.

The Model ID Did Not Change

This is the first thing to understand. The 0731 refresh is a post-training upgrade, not a new generation. You still call deepseek-v4-flash and you are served the latest build.

If your Hermes agent already runs DeepSeek V4 Flash, you are on the 0731 build already. No redeploy, no config edit, no model switch. The rest of this guide is for agents not on it yet.

What the Retraining Actually Did

DeepSeek published nine agent-focused benchmark results. The three with preview-build comparisons show how lopsided the gains were:

Benchmark	Preview build	0731 build
Terminal Bench 2.1	61.8	82.7
DeepSWE	7.3	54.4
DSBench-FullStack	37.0	68.7

For scale: DeepSeek's own V4-Pro-Preview scores 72.1 on Terminal Bench 2.1, so the cheap Flash model now runs ahead of the previous flagship preview on that test.

Read this honestly. These are DeepSeek's own numbers on agent-flavoured benchmarks, not independent evaluations, and a jump from 7.3 to 54.4 on DeepSWE says as much about how weak the preview build was on that test as about how strong 0731 is. Frontier models still lead on the hardest long-horizon work. What the 0731 build genuinely changes is the price-to-capability floor for routine agent loops — which is most of what a Hermes bot does.

Specs and Pricing

Property	Value
Model ID	`deepseek-v4-flash`
Context window	1M tokens
Max output	384K tokens
Direct API price	$0.0028 cache hit, $0.14 cache miss input, $0.28 output (per 1M)
Via OpenRouter	About $0.09 input, $0.18 output (per 1M)
Concurrency (direct)	2,500 — against 500 for V4 Pro

The cache-hit price is the number that matters for agents. A Hermes bot resends a growing conversation every turn, so most of its input tokens are repeats. At $0.0028 per million on a cache hit, a busy agent loop costs very little to keep running.

One thing to plan around: DeepSeek has said direct-API pricing will move to a peak/off-peak model where peak hours cost more. If you are budgeting a long-running agent, check the current DeepSeek pricing page rather than assuming today's rate is permanent.

Route 1: OpenRouter (Simplest)

Hermes treats OpenRouter as its default aggregator, so this is a two-command setup:

export OPENROUTER_API_KEY=sk-or-...

hermes inference set openrouter
hermes model set deepseek/deepseek-v4-flash

Or in the config file:

# /opt/data/config.yaml
inference:
  provider: openrouter
model:
  default: deepseek/deepseek-v4-flash

Note the deepseek/ prefix here — OpenRouter IDs carry a vendor namespace. This is the route most people should take.

Route 2: Direct DeepSeek API (Cheapest)

Going direct gets you the $0.0028 cache-hit price and the higher concurrency ceiling. DeepSeek's API is OpenAI-compatible, so it rides the custom slot:

# /opt/data/config.yaml
model:
  provider: custom
  base_url: https://api.deepseek.com/v1
  api_key: sk-...
  default: deepseek-v4-flash

Set provider explicitly. If you leave it empty, Hermes falls back to OpenRouter — and a bare deepseek-v4-flash with no vendor prefix is not an OpenRouter model ID, so the call fails. Note the model name is bare on this route and prefixed on the OpenRouter route; that difference trips people up.

Get a key at platform.deepseek.com.

Responses API and Codex Compatibility

The 0731 build natively supports the Responses API format and is advertised as compatible with Codex-style clients. If you have tooling written against that request shape, V4 Flash can sit behind it directly rather than through a translation layer.

On OpenClaw Launch

Managed Hermes instances route through OpenRouter already — pick DeepSeek V4 Flash from the model selector on your instance card, or switch from chat with /model deepseek/deepseek-v4-flash. To bill against your own key instead, see the Hermes BYOK guide.

Troubleshooting

402 from OpenRouter mentioning credits. The key is out of credit or capped, not a Hermes fault. Check your OpenRouter balance first.
Model not found on the direct route. You used the prefixed deepseek/deepseek-v4-flash against api.deepseek.com. Direct wants the bare ID.
Model not found on OpenRouter. The opposite mistake — the bare ID needs the deepseek/ prefix there.
Switching away later still hits DeepSeek. Clear provider, base_url and api_key when moving back to an OpenRouter model, or the stale custom slot keeps routing.
Costs higher than the cache-hit price suggests. Cache hits require repeated context. An agent that rewrites its whole prompt each turn never hits the cache.

FAQ

Do I need to update anything to get the 0731 build?

No. The model ID is unchanged, so any agent calling deepseek-v4-flash is served the latest build automatically.

Flash or Pro for a Hermes agent?

Flash for almost everything. It is roughly three times cheaper on output than V4 Pro, carries five times the concurrency, and after the 0731 retraining it matches or beats V4-Pro-Preview on several of the nine published benchmarks. Reach for Pro on the hardest long-horizon reasoning only.

How does it compare to Muse Spark 1.2?

Different price classes. Muse Spark 1.2 posts a higher Terminal-Bench score at $1.25 input; V4 Flash is roughly an order of magnitude cheaper and scores close behind. For high-volume routine agent work, Flash is the economical default.