ben

father, husband, optimist, entrepreneur, local-ai, crypto

[ben @benwvlshvm]

Let me tell you now that DeepSeek Harness is a joy to use.

When you put DeepSeek v4.1 Flash in it, it's basically AGI-at-home.

Day by day I am adding the features that I want/need and there is nothing really getting in the way of that happening...

Need an in-built notepad you can write ideas on during a session? Just ask DeepSeek via Creator Mode.

Need an in-built DGX Spark dashboard to monitor your compute on the go? Just ask DeepSeek via Creator Mode.

Need a full on, reverse engineered, Trello-style client and project management tool for your business to manage projects agentically? Just ask DeepSeek via Creator Mode.

After much testing and iterating with various local models and harnesses - for me personally - this feels the best to use day in and day out.

I'm slowly building all the pieces I need to be able to work effectively and efficiently in this new world we find ourselves in.

And the best part?

It's mine.

[ben @benwvlshvm]

Since I started posting and reading X way more I’ve slowly grown disgusted by Instagram… 🤢🤮

[ben @benwvlshvm]

@sudoingX There are places for both absolutely - owning your own compute much like owning your own place will be important - just another way to say fk you to the man!

[ben @benwvlshvm]

@zhngkn414984690 DSH first, Pi second.

[ben @benwvlshvm]

@CryptoBandit100 My entire business https://t.co/mn3EcE5zDf

Documentation, data analysis, scripts/code/automations - anything and everything now.

It’s all about the agent loop right and DeepSeek is trained to work within its harness. Feels OP having this aligned on my desk

[ben @benwvlshvm]

@Italianclownz Deepseek is so much more reliable and thorough for my own work. It’s just too good inside its own harness

[ben @benwvlshvm]

@yuhasbeentaken Yeah and with those 3 goals there isn’t anything better than deepseek inside of deepseek harness right now… even locally on 2 sparks im seeing 99% cache hit rates across many millions of tokens per session…. Feels damn good 😮‍💨😮‍💨

[ben @benwvlshvm]

@jayleaton It's consistent. Session after session... let's go!

[ben @benwvlshvm]

What a joy to use and watch it work...

I spent the last couple of weeks trying to get an internal tool working using GLM-5.3-Flash 4bpw with Hermes and it just couldn't get it done how I wanted

... I definitely take ownership of my lack of skill here so it's not a fair approximation of the quality of GLM+Hermes ...

BUT...

DeepSeek v4.1 Flash 1017 @ 2.8bpw with @jayleaton's quant/recipe got it done in 1 session.

Not only is the outcome better, it is so much more thorough - which for my work & business - is absolutely critical.

This is probably the first time since owning my Sparks that I genuinely feel like I have a capable teammate.

Just look at this stat line🤯🔥

[ben @benwvlshvm]

If you're in the knowledge work space - DeepSeek Harness is just toooooo good.. Keen to showcase some of the things I've been building here but

DeepSeek v4.1 Flash within DeepSeek Harness is just a joy.

Pretty much every single plugin I have asked for it has just built without a hitch.

All running locally via my 2x DGX Sparks... 😌

[ben @benwvlshvm]

@Sentdex Agreed. It’s really really good… few days in and not really looking back. All in deepseek!

[ben @benwvlshvm]

I can absolutely confirm this is true

[Jun Song @jun_song]

Local AI psychosis :

If you buy one DGX Spark,

you'll be adding a second one to your cart within a few weeks

[ben @benwvlshvm]

After running DeepSeek v4.1 Flash 1017 on my 2x Sparks over the past 24-48 hours with DeepSeek Harness, I am very excited for the near term future and what the opensource community can do here... 🫣

I do wonder whether they're looking at local agentic capabilities like Reverse Engineer Anything and that they see the writing on the wall - nothing closed is safe and you may as well get in front of it.

To be honest I have had a soft spot for DeepSeek ever since using it in China earlier this year... 😚

[DeepSeek @deepseek_ai]

🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!

Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.

🔗 https://t.co/JGAik1kgk0

1/4

[ben @benwvlshvm]

No problem! Keen to contribute where I can..

I had some issues with other components taking up RAM on Node 1. I disabled hermes and this gave me some more room.

But, straight from my agent:

WHAT IT TOOK — DeepSeek-V4.1-Flash-1017 (flash17 2.8bpw) on 2x DGX Spark, 4 sessions

Config (config/prod-zig-flash17.env) — diffs from the shipped profile:
TF_DSV41_CONTEXT=524288 # was 1048576 — THIS is the four-session change
TF_DSV41_SLOTS=4
TF_DSV41_POOL_TOKENS=2100000 # 4 x 524,288 = 2,097,152, so 4 full windows
TF_DSV41_PREFILL_CHUNK=2048 # was 4096
TF_DSV41_PREFILL_ROWS=2048 # was 4096
TF_DSV41_PF_4K=0 # drop the 4K prefill workspace
TF_DSV41_FLOOR_GIB=7 # was 5
TF_DSV41_GRAPHS_MAX=24 # default 48; I had seen it hold 96
TF_DSV41_VISION_MAX_IMAGES=32

On both hosts:
earlyoom -m 2,1 -r 60 # wedge insurance
no-swap cgroup slice for the rank containers (MemorySwapMax=0)
keep the serving node otherwise idle — no long agent turns on the head

Why the memory lines matter:
The low point is the PREFILL WORKSPACE, not the weights or the KV pool.
A 4K chunk allocates a ~2.27 GiB transient on BOTH ranks; at 2K it's ~1.1 GiB.

Result:
248k-token prompt OK — MemAvailable bottoms at 7.93 GiB (3.49 GiB when it OOM'd),
0 errors, 2.1M KV pool intact.

[ben @benwvlshvm]

What does 24 hours with DeepSeek V4.1 Flash 1017 look like? Here are some lessons from my first day running it locally with 2 DGX Sparks.

The first is that it's best used through DeepSeek Harness.

I have been using DeepSeek v4.1 flash through OpenRouter for a while now and running it locally here is nuts. It'll just keep going until the job is done.

(I have also fully migrated away from the clunky mess that Hermes is and am rebuilding my agentic workflows using DeepSeek...)

The second is that more often than not, sessions can run and run and run and run.

When the goal is short, it's super quick, when the goal needs to take time, it'll take its time. I am frequently hitting many millions of tokens of cached input at +98% cache hit rate with over 60 tok/s on average.

-- be careful what you prompt, because it could very well spend 30+ minutes going down a path you didn't intend.

The third is that it's extremely capable even at 2.8bpw (thanks to @jayleaton and his incredible work on this quant/recipe!). I have been jumping between local and API and honestly, I don't feel a huge difference in the outputs except for the obvious speed/concurrency etc.

The fourth is the way in which DeepSeek handles cache memory/storage. I did have to adjust a few recipe components and move context window to 524k tokens to accommodate for better concurrency across 4 sessions.

The fifth and last is; if you're running this locally, disable Agent Teams. While it is a really powerful plugin, it can cause a lot of context load which the Spark's struggle to handle.

This model is fun, and running it locally with DeepSeek Harness feels incredibly capable. It just feels good and it absolutely passes the vibe check.

The biggest thing for AI generally is - is it a vibe? Is it actually nice to talk to? The more we spend talking to these things, the more important it is that they're a delight to be around.

Huge progress from @deepseek_ai @DeepSeekHarness and again big thanks to @jayleaton for the incredible quant/recipe and @ashxhart for the TensorFold engine frameworks...

I feel each passing week my sparks grow in value and I continue to think maybe I should have bought 1 or 2 more 🫣🫣

[ben @benwvlshvm]

@landontgreen The vibes are definitely better with DeepSeek. It’s nice to talk to. It’s creative. Oh and also in DSH it’s an agentic workhorse. It just keeps on going on!

[ben @benwvlshvm]

@NullBurn GLM-5.3-Flash is great but I’ve noticed that GLM finishes quick and isn’t that thorough. V4.1-Flash on the other hand makes sure the job is done. My preference is DeepSeek after daily driving GLM the last few weeks

[ben @benwvlshvm]

@cdiamond That is for any 1 session. Across 4 sessions maybe I am at 30-40 Tok/s aggregate.

I see no drop off in speed even at long contexts

[ben @benwvlshvm]

@IterIntellectus I’d agree with this. A coworking space for me was game changing and recently going back to work from home (albeit in a new country) will be short lived for me 😅🙏

[ben @benwvlshvm]

@dummerspast39 Oh Wowee.. that’s immense. Keen to try this on my sparks ⚡️