DeepSeek-V4-Flash Update

You can’t treat Qwen3.6 like its fable, but if you prompt precisely and specifically it’s a great executor.

I actually found it refreshing to use more of my brain for once, and actually have to think deeper about what I’m trying to do, and how to build it.

https://github.com/anomalyco/opencode/issues/4276

Review is a skill, as in, a SKILL.md with a folder full of references:

- SKILL.md: https://gist.github.com/lionkor/161525be858d1d75db4c13c0f093...

- references/output-contract.md: https://gist.github.com/lionkor/8c68e33becef7a21f8408c7dc119...

- references/review-lenses.md: https://gist.github.com/lionkor/0a8b080fe45306213efddf3ebb75...

- references/review-workflow.md: https://gist.github.com/lionkor/d2d374b133ceb7e3660bd530ee72...

- references/section-rules.md: https://gist.github.com/lionkor/8a9e503adc7fd3697410cf021f27...

I'm aware that almost all of this is prompt voodoo, and there's no guarantee for the review to find anything or everything, but making it a dedicated skill and thoroughly observing the output thinking, tool calls, and result, lets me adjust these over time and fill the weak spots with even more prompting.

I use this skill by simply telling the agent something like "Review the changes on the current branch against origin/main, take special care with backwards-incompatible changes to the public API" or something like that.

I use `pi` (pi.dev) with a subagents extension, so that I can ask the agent to invoke a subagent to do the review, on work that the agent did.

For models, I use the highest possible reasoning on whatever model I feel like makes sense, usually this is GPT-5.5 or deepseek flash/pro, depending on the confidentiality of the codebase, on the highest reasoning always (for reviews).

I've also had success with a review checklist, though it doesn't produce an easy to parse (for humans) output: https://gist.github.com/lionkor/054ac2cf241e0765eee2383f0dba...

This is why my review skill mandates a very strict output contract. I need the output to be very easy to parse, and the output contract I've specified there does that.

In general I let <whatever the latest model of OpenAI's ChatGPT is> author and review SKILL.md and similar large prompts, usually with a ruleset like this, which is a 1600 line research artifact from a long GPT 5.5 "Pro" research session on prompt engineering: https://gist.github.com/lionkor/71498794d0a7d72173fc58766f25...

Does the review catch all issues? Not at all. Does it catch, usually more than one, important issue, across large changesets? Absolutely, and that's the point! :)

Feel free to ask me any questions, I'm also happy to share more about my setup via email or add you or anyone else to my private repos with more of these.

For any work with protected intellectual property, I use other providers, for the contractual guarantees, but I think it would be silly to think that OpenAI or Anthropic are not training on literally all data they get. How could you ever tell if they did? They can just claim the data was mislabelled, or ignore the accusations. If you have serious IP, use only local models.

The main difference is that deepseek is bad at prose, and Codex models are much more eager to use tools provided to them (which is usually fine, they often make like 3 todos via tool calls for a simple one step task which is silly though), and deepseek in general benefits from good instructions more than GPT models maybe do.

Deepseek becomes much better if you give it tools for asking clarifying questions, doing self-review with subagents, and so on. The more tools it uses, the better the signal-to-noise ratio in its context, and the more consistent the output.

You must still review 100% of the code as if it's trying to sell you insurance.

Then you, dakolli - out of generosity and minute-to-minute devotion to enlightenment - sacrifice time from your busy day to sit down (though perhaps that's been painful lately?) or stand up with your phone - and offer a profound, deeply thought-out counterpoint in the following form (and I'll paraphrase):

"Nah mate, it's shit. All LLMs are shit."

Usually unreviewed code only gets committed if I really don't care, like for one-off scripts, which I sandbox with github.com/lionkor/sbh or run as an unprivileged user.

To be more constructive, what I did (as an experiencd SWE but a complete noob to agentic coding): went to pi.dev's extension marketplace and looked into all the new shiny stuff. Subagents, mcps, context and memory optimizers, skills. Using the most popular ones (not necessarily the best ones)

It was like 15years ago learning the new mindset of vim (and spending a ton of time to customize it to my workflow). My understanding is that Claude and opencode doesn't give you this flexibility.

Learning all these stuff drove me to also set up openwebui, and it was such a successful private project that I implemented it at work (with jira/confluence/bazel query access) and management said "we need this by tomorrow".

I believe the models matter not that much anymore. The "harness" does. (unless you just want to vibe code. Thebn, throw crap at fable and call it a day)

https://github.com/gitsense/pi-brains/tree/staging

The README is being worked on but the three videos should give you a good sense of what it can do. Pi is also what makes what I will demo in

https://github.com/gitsense/chat/tree/update-readme

possible. Since Pi exposes so much, it is very easy to build advanced tooling around it to help easily grok hundreds of tool calls to help you understand what they agent knows and what it has tried.

Lots of sensible defaults and good tweaks/settings you just don't worry about.

Last night it single handedly implemented a feature after a grilling session, and came back with the red-yellow-green risk assessment points that I mostly saw with anthropic models. I had to check if I'm using the right model, but it was DS4Flash.

So maybe I was using it already?

* DS4 Flash: 82.7

* GPT 5.6 Luna: 75.7

For reference, that puts it on the third spot behind GPT 5.5 and Fable 5. For some reason GPT 5.6 Sol is not showing in the leaderboard. If it did, then DS4 Flash was number four.

The thing is, even if Luna is better in DeepSWE and has the 80% discount. DeepSeek is still cheaper.

--------------

DeepSeek V4 | Flash GPT-5.6 Luna (New)

--------------

Input (Cache Hit) $0.0028 $0.02

Input (Cache Miss) $0.14 $0.20

Output $0.28 $1.20

--------------

Both Luna and Flash are heavy on the reasoning > output. And the cache hitrate + prices also matter.

Reality is, you can not go wrong with Luna or Flash at those prices. And remember, DeepSeek V4 Pro is still in the rafters, what is ironically closer to Luna's new price.

In terms of open research, China has absolutely overtaken the US.

US or American, don't trust anyone, these are open weight models. Host them yourself if you feel strongly about privacy (as you should). I honestly don't know of any OSS open weight models from the US labs as good as Kimi K3 or Deepseek V4 though.

There is no telling how much worse US will get as they start to lose more and more internationally and grow the wealth inequality locally.

It is easy to be liberalist and cool when winning but it doesn’t mean there isn’t a racist, uneducated and violent majority

They release the models back for free.

Undermining out entire economy by subsidizing the release of DIY versions of our main economic drive sounds like a huge win for China.

Sarcasm aside - if the community can't get their shit together to continue improving what is currently public, well - we don't deserve free stuff and open weights.

Dual Strix Halo is much slower and current Macs with 256GB are both slower and more expensive (Mac Studio M3 Ultra 256GB around 12000€).

To get something faster than the two Sparks you'd need to spend more than $22000 for a server with 2x RTX Pro 6000 at $10000 each.

Beyond that you could get 2x AMD MI350P.

You can also do 2x 6kPRO in a workstation, for ~20k.

I guess using a ZDR provider is good enough for now.

  Excuse
  "Too simple to test"
  
  Reality
  Simple code breaks. Test takes 30 seconds.

https://opencode.ai/docs/go/#usage-limits

Previously, OpenCode Go had higher API prices for some models, but now they lowered the API price and simultaneously reduced the allowance.

Providers' cache hit stats are available to consult, and only 1-2 of them behave properly if I remember correctly, zero if you request providers that don't store and train on sessions.

The `-0731` style suffix is worse compared to a proper version bump like V4.1.

Some providers like OpenRouter now call it `deepseek-v4-flash-0731`, but even in places like here on HackerNews people say things like "Sonnet is better than DeepSeek" without specifying a version or a reasoning effort, certainly no one will mention that `-0731` suffix when talking about DeepSeek V4 Flash.

> 78.4

The real score is always the official benchmark.

We need to see later if DS4 flash 0731 is going to maintain the score but we need to look at the official benchmarks.

Already seen a PR for DeepSWE to update the benchmark with 0731, so we can verify claimed vs official.

They're literally comparing the previous version of the same model with the new one. It's based on the same architecture, same pre-trained model, just different post-training. It doesn't get more apples to apples than this.

The performance changes are so big with the right harness that is makes sense to engineer the harness and fine-tune the model to one another from the start.