I Don't Need Frontier Models Anymore (Qwen 3.8 27B + DeepSeek Harness) — Transcript

Full transcript

1. 0:01When 3.8 27 billion is out, okay, this
2. 0:04is the model that I was waiting for.
3. 0:07I've been using the previous version,
4. 0:09the 3.6 27 billion, for quite a while
5. 0:12and it's become my daily drive. I do 80%
6. 0:15of the work with that model.
7. 0:17I like it a lot. I bought a new computer
8. 0:20just to run that model.
9. 0:22I now run it on the 5090 and it's
10. 0:24performing amazing. So, you can imagine
11. 0:27how much I was waiting for this new
12. 0:29version.
13. 0:31And unfortunately, I have mixed feeling
14. 0:34about it.
15. 0:35Now,
16. 0:36I want to explain because this is really
17. 0:38important. There is something extremely
18. 0:40important to understand, but also I have
19. 0:42an insane good news about this model.
20. 0:45This is what when released and this is
21. 0:48the reason why a lot of people are
22. 0:49comparing it to Opus 4.6. I personally
23. 0:53I'm not much interested in this model.
24. 0:55I'm more interested to understand what
25. 0:57I'm gaining from what I had before to
26. 0:59what I can run today.
27. 1:01And I can see that in some scenario, the
28. 1:04gap is not that big,
29. 1:06but other scenarios like here, agent
30. 1:08decoding, the gap is huge from 13.3 to
31. 1:1242.2, it's massive. And software
32. 1:15engineer for 49.3 to 79, this is another
33. 1:20massive step forward. On frontier agent
34. 1:23task,
35. 1:25from 10.6 to 20, this is 100%
36. 1:28improvement. The gap sometimes are
37. 1:32pretty noticeable. So, why I felt this
38. 1:35mixed feeling when I I tested myself?
39. 1:38Because
40. 1:39I tested the Qwen 3.6 27 billion
41. 1:44Q6
42. 1:45with the Qwen 3.8 27 billion Q5. In my
43. 1:50test, bug hunting, the Q6 found one more
44. 1:54bug than and Q5, even though the Q5 was
45. 1:57the newer version. So, I stopped the
46. 2:00test and I said, "Okay, this is
47. 2:02This is disappointing for
48. 2:04us, but let's compare them with the same
49. 2:06quantization." So, I also installed the
50. 2:08Qwen 3.8 27 billion Q6, then I could see
51. 2:12the difference. But, the difference was
52. 2:14still not that big as much as I was
53. 2:17hoping for.
54. 2:19So, there is probably it's hard to tell,
55. 2:21but let's say a 30% improvement, which
56. 2:24is noticeable, but not another level.
57. 2:27So, I was expecting much more. This is
58. 2:29because Luna itself could do such a
59. 2:33great job when improving reasoning. So,
60. 2:36here is GPT 5.6 Luna non-reasoning 27.
61. 2:41When we start increase reasoning,
62. 2:44go to 34, then
63. 2:4639,
64. 2:48then
65. 2:4947, and then GPT 5.6 Luna max 52. The
66. 2:55jump is insane. It's more than double
67. 2:58the capability.
68. 3:00So, let's look at
69. 3:02Qwen 3.6 27 billion non-reasoning 31.
70. 3:06When does reasoning go to 38? I was
71. 3:09expecting, okay, the new model is going
72. 3:11to be better as a model, but plus then
73. 3:13going to increase the reasoning, so it
74. 3:15should push us towards this kind of
75. 3:18level. We are probably in the 40, but I
76. 3:21was hoping to get actually up to in the
77. 3:2350s. That is reality. Considering that
78. 3:27when non-reasoning 31 and Luna
79. 3:30non-reasoning is 27. So, potentially, we
80. 3:34could have done much more than 50 and
81. 3:37reach, you know, this kind of level.
82. 3:40The DeepSeek version 4 Pro.
83. 3:43That would be amazing. The GLM 5.2 max.
84. 3:46Okay, that would be an amazing
85. 3:48achievement, but we didn't. So, probably
86. 3:51I was expecting too much.
87. 3:54But let me tell you the good news here.
88. 3:58The good news is this model is capable.
89. 4:02But
90. 4:04intelligence is not just about the
91. 4:06model.
92. 4:07Is the harness around it.
93. 4:10Before I explain why
94. 4:12I'm talking about harness now, it
95. 4:13because I found that the killer combo
96. 4:16between this model and a harness. Until
97. 4:20now I was using Hermes and Open Code.
98. 4:24And while Hermes I find it that is a
99. 4:27fantastic system.
100. 4:30Recently, when they start pushing their
101. 4:33own subscription, I felt like
102. 4:36yeah, that is not cool anymore. I felt
103. 4:39like a client.
104. 4:40I'm a guest within their system, and
105. 4:43this is the reason why I'm building with
106. 4:45my community Resonate OS, so we can
107. 4:48co-own it and it's ours.
108. 4:50The moment they started pushing their
109. 4:53subscription, the moment that people are
110. 4:54starting asking me is do I need to
111. 4:57subscribe to use Hermes?
112. 4:59I could see that
113. 5:01Hermes just took the wrong path.
114. 5:05Okay, and therefore I was already open
115. 5:07with the idea I need to look to
116. 5:09something different.
117. 5:11With Resonate OS, I was thinking, okay,
118. 5:13for us everything needs to be an add-on,
119. 5:16okay? All the elements need to be
120. 5:19possible for the community to add to the
121. 5:21system, so they can customize, they can
122. 5:22change it, they can replace part, they
123. 5:24can replace the memory, they can replace
124. 5:26the main agent, they can replace all
125. 5:28those parts, but at least we have a
126. 5:30system that we can work together and
127. 5:32build it and customize together.
128. 5:35So, at one point I came across randomly
129. 5:38to this new harness that came out
130. 5:41recently. I think less than a a week ago
131. 5:44Deep Seek released Deep Seek harness.
132. 5:48Now, I believe that this is the best
133. 5:51combo for when 3.8 27 billion.
134. 5:56I started using this harness. I'm going
135. 5:58to explain a little bit, even though I
136. 5:59will probably need to do a dedicated
137. 6:02video to this harness because it does
138. 6:04something that no other harnesses do.
139. 6:07What happened? I started using it
140. 6:08together with this new version of Quen
141. 6:11and everything I wanted to build so far,
142. 6:14it built it. So, in the past, I used to
143. 6:17use Quen 3.6 27 billion, let's say 80%
144. 6:21of the time. Sometimes, it would find a
145. 6:24problem that was not solvable. I had to
146. 6:26change model and use one of those
147. 6:29frontier model. Now, with this Deep Seek
148. 6:32harness and the new 3.8 27 billion, I
149. 6:35never
150. 6:37had a problem to swap to a different
151. 6:39model.
152. 6:40So, what is happening here? When I went
153. 6:43to this website, I read everything is a
154. 6:45plugin. So, in my world, everything is
155. 6:48an add-on, for them, everything is a
156. 6:50plugin. It resonated and said, "Hmm, I
157. 6:53need to look into this thing."
158. 6:55I was coming from
159. 6:57not totally sure about Hermes anymore.
160. 7:00We are still building with our community
161. 7:02resident of us.
162. 7:03I see this that says everything is a
163. 7:06plugin.
164. 7:07Total alignment. I installed it.
165. 7:09Actually, I didn't even install, I asked
166. 7:11Hermes to install it for me and
167. 7:13configure it with the new model
168. 7:16and it's amazing. So, now, this is
169. 7:20how it looks like, okay?
170. 7:22And I can't show you while it's running
171. 7:25because while it's running, I can't
172. 7:27record the screen because memory
173. 7:29management of the file format that I'm
174. 7:32recording. What I can tell you here is
175. 7:36that is just amazing. It's not a
176. 7:39finished product, okay? Okay, so they
177. 7:41call it
178. 7:42developer preview. So it's not finished.
179. 7:45It's there is a lot of things maybe I'm
180. 7:47missing things that are not perfect. So
181. 7:49sometimes it doesn't do compaction in
182. 7:52the right moment.
183. 7:54And therefore it's a kind of a crash the
184. 7:56system. Nothing serious because you just
185. 7:58say resume working and it start working
186. 8:01without any problems. But nevertheless
187. 8:03it's still a preview. But what it did
188. 8:06they change how the context window is
189. 8:09managed.
190. 8:10What is happening here? It's
191. 8:12unbelievable.
192. 8:13There is a lot of new elements inside.
193. 8:17So it takes track of everything that has
194. 8:20been done.
195. 8:21And require for whatever system that
196. 8:24they use less context window. I can't
197. 8:27explain because it's confusing to me so
198. 8:29I need to read and understand it better.
199. 8:31But what I know it's this 131,000 tokens
200. 8:35that I have available for this model on
201. 8:37my 5090.
202. 8:39They just last longer. The compaction is
203. 8:43so much more solid. And this system is
204. 8:46being chatting for so long building for
205. 8:49so long that is incredible. So if you
206. 8:52see I don't know if you can see but here
207. 8:54it had an input of 38 million tokens.
208. 8:59Still going strong. What did I build
209. 9:02with this? I build this the augmentor.
210. 9:06This is my agent. Okay, this is how I
211. 9:07call my agent. This is part of resonant
212. 9:10OS is also included in resonant OS. The
213. 9:12ability augmentor that is powered by
214. 9:15this deep seek harness.
215. 9:17What it does it's it takes total control
216. 9:20of the browser and and just go around
217. 9:24and click and do things for me. So what
218. 9:27I ask here was to check out of my ladies
219. 9:31videos to see the comments and suggest a
220. 9:35few videos idea.
221. 9:37So that was a way to send
222. 9:39these agents into the browser, click
223. 9:42around, and grab information.
224. 9:46So, here is what came out. It checked
225. 9:49four five different videos.
226. 9:52It saw some of the comments, and then
227. 9:54come out with some ideas.
228. 9:56Now, it doesn't matter the ideas because
229. 9:58it's honestly I don't from what I see
230. 10:01maybe are valuable, maybe not, but
231. 10:03doesn't matter because it didn't have
232. 10:05the instruction of what is this channel
233. 10:07is about and so on. So, at the moment it
234. 10:08doesn't have much context. It just went
235. 10:11there and grabbed information. But, what
236. 10:13is important is what was capable of
237. 10:16doing. I built the all thing with this
238. 10:19Qwen 3.8.
239. 10:21It works completely fine. Yeah, there is
240. 10:24a a little things that need to be
241. 10:25improved.
242. 10:26And the most important part, I never had
243. 10:29the need to use a bigger model.
244. 10:32Something that always happened before
245. 10:34when was doing something as complex as
246. 10:37this kind of tasks. So, we have another
247. 10:40deep seek moment, but this time the deep
248. 10:42seek moment is not on the model, is on
249. 10:44the harness, and is done by deep seek.
250. 10:47Now, I'm extremely curious to see what
251. 10:49this harness can do even with a more
252. 10:52powerful model. If you've been following
253. 10:54this channel, you know that I'm really
254. 10:56pro local AI and so on.
255. 10:59I'm for the AI sovereignty, meaning you
256. 11:02need to own the hardware and
257. 11:04infrastructure so you can trust and
258. 11:07control everything. This is the journey,
259. 11:09but it's not first of all it's not for
260. 11:11everybody, but also is is a journey,
261. 11:13it's not an on off.
262. 11:15What I can see with this deep seek
263. 11:16harness that the journey is getting
264. 11:18better and better because now we have
265. 11:20this harness that has MIT license,
266. 11:24meaning you can use it in your project.
267. 11:27You can use it yourself, implement,
268. 11:29change it, do whatever, and use for
269. 11:31commercial. So, what I'm suggesting
270. 11:33here,
271. 11:34if you are watching this video, probably
272. 11:36is because you have access to this
273. 11:38model.
274. 11:40But, even if you don't, try this uh
275. 11:42DeepSqueak Harness.
276. 11:44If you're one of those lucky people that
277. 11:46own a
278. 11:48two DJX Spark
279. 11:50or equivalent,
280. 11:51try DeepSqueak version full flash,
281. 11:54because can run on two of them with this
282. 11:57harness, that probably they could work
283. 12:00amazingly together. I'm at the moment of
284. 12:02after a few months of uh
285. 12:05of struggling with this AI,
286. 12:07I'm excited again. I think it's a
287. 12:10similar moment of when Open Claw came
288. 12:13out, which was called uh
289. 12:15Clawbot, I think. Can't even remember
290. 12:17anymore.
291. 12:19So, when it came out, it just was a
292. 12:23massive push forward, because everybody
293. 12:25finally had this agent that can do a
294. 12:27lot. They control your computer, work
295. 12:29for you, and so on. So, now I think what
296. 12:33DeepSqueak did with their harness is a
297. 12:36similar jump.
298. 12:38Allows this model to do so much more,
299. 12:42which is incredible.
300. 12:43I will need to test this for longer, of
301. 12:45course. I can't say,
302. 12:47you know, that is the final solution,
303. 12:50but at moment, this is what I feel. I
304. 12:52feel like I never needed a more powerful
305. 12:56model than this when I use this harness.
306. 12:59Something that I I can't say the same
307. 13:01with uh Hermes or Open Code.
308. 13:05I think it's a killer combo. I need to
309. 13:07understand this DeepSqueak harness even
310. 13:09more. At moment, it is not fully
311. 13:11finished. The way it works it's uh
312. 13:14surprisingly well. So, if I were you,
313. 13:17first of all, go to this DeepSqueak.com
314. 13:21harness and uh and read it through. Like
315. 13:24I said, I will do a new video, but they
316. 13:26are definitely managing the things
317. 13:28differently. And now, what is possible
318. 13:30to achieve with this harness and the
319. 13:32Qwen 3.8 27 billion is impressive.
320. 13:37This is going to be my new way of
321. 13:40working from now on. I will need to
322. 13:43migrate and learn how to do it from
323. 13:45Hermes to this model.
324. 13:48Again, the good thing that is happening
325. 13:50that is I basically already built this
326. 13:53agent. This agent they are going to
327. 13:55mentor.
328. 13:56This can be a an add-on for Resonant OS
329. 13:59and again shows the importance of
330. 14:01working in add-ons or plugins like they
331. 14:05call it because now it's easy to
332. 14:07implement new things.
333. 14:10And what I can even believe is that I
334. 14:14built this
335. 14:15100% with my local AI.
336. 14:19Not needing a more powerful system.
337. 14:23Everything that I asked to do has been
338. 14:25done. Not getting lost and so on. One
339. 14:29thing that I want to say about the
340. 14:30psychic harness connected together with
341. 14:33this Qwen 3.8 27 billion
342. 14:36is they use a lot of token.
343. 14:39They use a lot of tokens.
344. 14:41But,
345. 14:44the result is there. We know that the
346. 14:46direction was increase the thinking time
347. 14:48and you you will perform better. Here is
348. 14:51happening and is incredible. Let's look
349. 14:54at the what kind of hardware you need to
350. 14:56run this model. In theory, according to
351. 14:58Qwen, you can run it on 12 GB of RAM.
352. 15:03Ideally, a 48 GB of RAM. And I would
353. 15:06agree on the 48 GB of RAM because what
354. 15:09you gain is actually a larger context
355. 15:11window.
356. 15:12I have the 32 GB of RAM. Okay, a Q6
357. 15:16and 25 GB and look, I only lose a 0.2%
358. 15:23from the Q8. So, the compromise is
359. 15:26pretty acceptable. With the Q5, you lose
360. 15:29a little bit. It's you lose 1.4.
361. 15:32At 1.4, you say, "Okay, that is not
362. 15:34much." And I mean, from this one, but
363. 15:37you lose a 1.6
364. 15:40from the top. It's still not that much,
365. 15:42but from my benchmark that I run
366. 15:45earlier,
367. 15:47the bug hunting, the Q6 could find one
368. 15:51more bug than the Q5.
369. 15:55So, it is a little bit, you know,
370. 15:57annoying, this thing. So, here's my
371. 16:00suggestion.
372. 16:01If you're looking for a bigger contest
373. 16:03window, Q5 is fine. If you need to do
374. 16:06some specific work that is more
375. 16:08advanced,
376. 16:10of course, you need to move into the Q6,
377. 16:13ideally.
378. 16:14If you have the 32 GB of RAM. I'm
379. 16:17explaining this because it's important
380. 16:20to understand what are the limits. When
381. 16:22you go into the 16 GB of RAM, you start
382. 16:25to lose quite a bit. At 12 GB of RAM,
383. 16:29you lose a lot.
384. 16:32So, while I never tried these two
385. 16:35models, okay, I never tried them, I
386. 16:38can't say how much you're losing from
387. 16:40the model. This is something that you
388. 16:41need to test yourself. But knowing that
389. 16:43the minimum minimum requirement is 12,
390. 16:46ideally, you want to be in the 24, 32,
391. 16:5048 if you can do it. So, now, another
392. 16:53element you need to consider is the size
393. 16:55of the memory that is required to run
394. 16:57this model.
395. 16:58But the speed at which the model runs is
396. 17:02connected to the bandwidth.
397. 17:04So, I just listed here, I used these
398. 17:07perplexities to do this research on the
399. 17:1032 GB of RAM and 24 GB of RAM. So, here,
400. 17:14for the 32 GB of RAM, the best one is
401. 17:17the 1590, okay? This has a bandwidth
402. 17:20that is insane, 1,792
403. 17:23GB per second. That means the this model
404. 17:27is working at around 100 and 10 token
405. 17:30per second on my hardware. It's doing
406. 17:32insanely well.
407. 17:34So,
408. 17:35when you go on this one,
409. 17:38you can see that the AMD is doing 640.
410. 17:43This is almost a third of the speed. So,
411. 17:45you have an idea of what you can expect.
412. 17:49Now, on the 24 GB of VRAM, we have a
413. 17:53different kind of opportunity.
414. 17:551,000 GB per second, which is pretty
415. 17:58good. 3090 is the same. The good thing
416. 18:01with 3090 that you can find them second
417. 18:03hand much cheaper. This is the 3090
418. 18:05idle. So, the 3090 normal 3090 is here,
419. 18:09which is
420. 18:11basically
421. 18:12not not a big of a difference, okay? It
422. 18:13doesn't really matter.
423. 18:15But, from AMD, we have this card,
424. 18:19which is still in commerce, that you can
425. 18:21buy and it's not as expensive as those
426. 18:254090. And the speed is almost there.
427. 18:28Now, those are 24 GB of RAM. So, if I
428. 18:31were in you and if you can afford, I
429. 18:34mean, two of these together, they give
430. 18:36you 48 GB of RAM, which again give us
431. 18:41this 48 GB of RAM. I would not run the
432. 18:44Q8. I would go for the Q6, but with a
433. 18:46much bigger context window. Keep in mind
434. 18:49that on my 32 GB of RAM, I can have
435. 18:52131,000
436. 18:54token context window. With 48, you
437. 18:57probably, I guess, you could have the
438. 18:59the total 260 something context window,
439. 19:03which is a nice plus.
440. 19:05Remember, Q8 is going to run slower than
441. 19:08the the Q6, the Q6 slower than the Q5,
442. 19:11and so on.
443. 19:12Bandwidth, it is something to consider.
444. 19:14At the moment, if you want to buy new,
445. 19:16this is really good. If you're looking
446. 19:18for second hand, this is pretty good.
447. 19:21One more thing, this model can run on a
448. 19:24DGX Spark and it runs, depending on the
449. 19:26quantization, at around 20 30 token per
450. 19:29second. At 30 token per second, in
451. 19:32reality, you have a system that works,
452. 19:34is usable, and I will probably run it
453. 19:36and test it on my GX 10.
454. 19:39This is the ASUS version of the DGX
455. 19:41Spark. It's 128 GB of RAM, so I can run
456. 19:44it with really large context window with
457. 19:47multiple instances. So, that would be
458. 19:49the advantage.
459. 19:51Anyway, that's is what I wanted to share
460. 19:53for today.
461. 19:54To recap,
462. 19:56a model that is not as good as I wanted.
463. 19:58It didn't do the jump that I was
464. 20:00expecting and wanting. Probably we're
465. 20:02going to see it when they release the
466. 20:03version four of it. But, combining this
467. 20:07with Deep Seek harness, I saw something
468. 20:11that I never saw before. It's early days
469. 20:13to say, but so far, in the last 24
470. 20:15hours, I use it a lot.
471. 20:17And I never felt I needed to use a
472. 20:22bigger model. Never once.
473. 20:24[clears throat]
474. 20:25That make me think that things are
475. 20:27changing and locally I is gaining more
476. 20:31and more value
477. 20:33every single day.
478. 20:34If you're interested in this kind of
479. 20:35content and you want to join our
480. 20:37community, first like and subscribe. In
481. 20:39the description, you'll find the link to
482. 20:41our Discord channel and there is a deep
483. 20:44dive article. I will go deeper in
484. 20:47everything that I'm talking inside this
485. 20:49video. I will add all the links that I
486. 20:52discussed over here and if possible, I
487. 20:54will see how it goes, but I might also
488. 20:57share
489. 20:58this augmentor power by this DSH. This
490. 21:02is an extension for the browser and by
491. 21:06adding it, you can have this sidecar.
492. 21:10Which, honestly, is amazing. It is
493. 21:13actually amazing.
494. 21:15I say if I can, because what is missing
495. 21:18before I can share it is a way for you
496. 21:21to add your own model. At the moment,
497. 21:24there's no
498. 21:25configuration system where you can add
499. 21:28your API or local model or something.
500. 21:31Even though through the code, you could
501. 21:33be able to ask to your agents to add the
502. 21:36model, I want to make it through the
503. 21:38interface so everybody can just add
504. 21:40their, you know, details and run it
505. 21:43locally or for any other cloud provider.
506. 21:47So, if I can manage to do this in time,
507. 21:50I will link inside the Substack article
508. 21:53the code for this
509. 21:55agent. If not, will be shared in the in
510. 22:00the near future. Thanks for staying
511. 22:01until the end. Again, like and
512. 22:03subscribe. Watch the next video. Join
513. 22:05the Discord. And ciao.