Pinned Tweet
Ripple-65B solved 8 of 15 hard coding tasks. Qwen3.8 solved none. 👇
I tested Ripple-65B-Preview against Qwen3.8-Flash-Next and Qwen3.8-27B on 75 brand-new tasks. A script generates every task from a random seed, so none of them can be in anyone's training data. Code checks Ripple passed 58 of 75 (77%). Flash-Next passed 47 (63%) and the 27B passed 48 (64%).
The real story is code. One task family asks the model to write a Python interpreter for a made-up expression language, graded on hidden tests. Ripple passed 8 of 15. Every Qwen run passed zero. That gap is solid (paired test p = 0.008 against each Qwen model).
The Qwen models didn't fail on wrong code. They ran out of room: every single reply hit the 32k-token cap. Ripple's passing answers took 21k to 31k tokens, so it was tight for Ripple too (it hit the cap 4 times).
On the other four families (ledger maths, loan interest, scheduling, JSON rules), none of the models are clearly ahead.
I also ran the 27B on my MacBook Pro and on the Spark with the same 4-bit weights, to check that Metal vs CUDA wasn't skewing things. 48 vs 42 is within noise.
What's next: first, some big performance improvements for Ripple on TensorFold. Then I'll dig into the code gap with other languages and longer context, to see whether the Qwen models get there with more room. Assuming Qwen 4 doesn't drop first 😅
Setup: one try each, thinking on, T 1.0, top-p 0.95, top-k 20. 15 tasks a family, so the error bars are wide. Ripple ran on vLLM, the Qwen models on TensorFold.
@RicursiveAI
I tested Ripple-65B-Preview against Qwen3.8-Flash-Next and Qwen3.8-27B on 75 brand-new tasks. A script generates every task from a random seed, so none of them can be in anyone's training data. Code checks Ripple passed 58 of 75 (77%). Flash-Next passed 47 (63%) and the 27B passed 48 (64%).
The real story is code. One task family asks the model to write a Python interpreter for a made-up expression language, graded on hidden tests. Ripple passed 8 of 15. Every Qwen run passed zero. That gap is solid (paired test p = 0.008 against each Qwen model).
The Qwen models didn't fail on wrong code. They ran out of room: every single reply hit the 32k-token cap. Ripple's passing answers took 21k to 31k tokens, so it was tight for Ripple too (it hit the cap 4 times).
On the other four families (ledger maths, loan interest, scheduling, JSON rules), none of the models are clearly ahead.
I also ran the 27B on my MacBook Pro and on the Spark with the same 4-bit weights, to check that Metal vs CUDA wasn't skewing things. 48 vs 42 is within noise.
What's next: first, some big performance improvements for Ripple on TensorFold. Then I'll dig into the code gap with other languages and longer context, to see whether the Qwen models get there with more room. Assuming Qwen 4 doesn't drop first 😅
Setup: one try each, thinking on, T 1.0, top-p 0.95, top-k 20. 15 tasks a family, so the error bars are wide. Ripple ran on vLLM, the Qwen models on TensorFold.
@RicursiveAI
1
0
0
166
Own your compute now or pay forever.
They switched from Grok-4.7 to Opus-5.5 a few weeks ago.
They started losing too much money, so they rug pulled and nerfed the quota last week.
Then, people noticed their $300/m SuperGrok Heavy subscriptions were running out after 8 hours instead of lasting a full week, so now they've "restored" quota, but switched to some cheapo ChatGPT-6 model.
Cloud AI disappearing is inevitable. But the endless thrash for customers is immensely frustrating.
0
1
1
5
BREAKING:
Elon Musk announces substantial improvements for Grok Bot usage and performance. 👏👏👏
It is unclear if there will be any reset for existing users who are locked out at 100% usage.
Elon Musk announces substantial improvements for Grok Bot usage and performance. 👏👏👏
It is unclear if there will be any reset for existing users who are locked out at 100% usage.
Just did a round table review with the Grok @bot engineering team.
Fixes are being implemented overnight and you should notice a dramatic improvement by tomorrow. If not, reply to this post and we will address any remaining issues right away.
0
0
2
118
Good morning from Canberra, Australia!
Two weeks of posting consistently starting with zero followers, and my Qwen 4 post just hit 9.4k impressions.
I love talking about local AI, open-weight models, and what actually runs on real hardware. I'll be testing Qwen 4 on my DGX y it lands.
If you're building for the future, let's connect 🤝
Two weeks of posting consistently starting with zero followers, and my Qwen 4 post just hit 9.4k impressions.
I love talking about local AI, open-weight models, and what actually runs on real hardware. I'll be testing Qwen 4 on my DGX y it lands.
If you're building for the future, let's connect 🤝
3
0
3
329
BREAKING
X user @follistatindev claims to have early access to Qwen 4 Flash and 27B.
The crazy part?
Qwen 4 27B performs similar to Claude 5.5 Sonnet.
Qwen 4 Flash performs better than Opus 5.5.
If true, this would be a genuine revolution in local AI.
Will Qwen 4 be enough to local AI?
X user @follistatindev claims to have early access to Qwen 4 Flash and 27B.
The crazy part?
Qwen 4 27B performs similar to Claude 5.5 Sonnet.
Qwen 4 Flash performs better than Opus 5.5.
If true, this would be a genuine revolution in local AI.
Will Qwen 4 be enough to local AI?
Qwen-4-Flash-Early (codename hawk) added to GameFeel Bench!
(and audio score was calculated wrong for qwen-4-27b, a genuine mistake that was fixed)
5
3
104
10,663
Interesting take here from @grok on automation of data analyst roles.
Not what I expected.
Do you agree?
Not what I expected.
Do you agree?
Never fully, in any realistic timeframe.
AI already handles 40-60% of the mechanical work (SQL, cleaning, basic reports, dashboards). That share will hit 80%+ for routine tasks within 5 years. Junior report-generator roles are already being cut or skipped.
What remains—framing ambiguous questions, domain judgment, validating AI output, owning decisions, explaining trade-offs to stakeholders—is not on a clear path to automation. That core stays human for the foreseeable future (10-20+ years).
The job is transforming fast. Adapt or get replaced by people who do.
1
0
3
199
Thought I'd check out the local @NVIDIAGeForce 5090 market.
Holy guacamole.
These were $7,500 a couple of months ago and everyone thought that was insane 🤣
Guess I'm sticking to the M5 Pro and DGX Spark!
Holy guacamole.
These were $7,500 a couple of months ago and everyone thought that was insane 🤣
Guess I'm sticking to the M5 Pro and DGX Spark!
0
0
4
126
Ripple-65B-Preview's DFlash2 drafter is now adapted to TensorFold's native CUDA engine, on one DGX Spark (GB10) 👇
Model PR: github.com/ashhart/Tensor…
Drafter PR: github.com/ashhart/Tensor…
The engine reads the NVFP4 release and the publisher's DFlash2 drafter as published. lain replies bit for bit (22 of 22, and 0 differences in 144 checks).
One stream, plain → drafted:
• Code: 5.2 → 26.1 tok/s (5.0x)
• Prose: 4.4 → 10.0 tok/s (2.3x)
Eight streams at once, in total:
• Code: 37.5 → 49.5 tok/s (1.32x)
• Prose: 32.6 → 40.3 tok/s (1.24x)
The smaller eight-stream gain is because each stream only gets about 1.5 drafts a round, versus about 7 to 10 on one stream. Which leaves me thinking there is a lot more power to unlock in this model.
Still, some extraordinary results from a single DGX Spark.
Prompts: about 520 tok/s profile; first token in 3.9 s for 2,048 tokens. Against the publisher's model code run in float64: top-1 agreement 98.7%, mean KL 0.0007.
Next up: quality evaluations against Flash Next and 27B, then longer drafts when several streams share a round, and faster prompt kernels.
GB10 only. 256-token replies, median of 3, greedy. Code is a raw text completion; prose is a chat reply.
@ashxhart @RicursiveAI @MiaAI_lab @bertholomusai
Model PR: github.com/ashhart/Tensor…
Drafter PR: github.com/ashhart/Tensor…
The engine reads the NVFP4 release and the publisher's DFlash2 drafter as published. lain replies bit for bit (22 of 22, and 0 differences in 144 checks).
One stream, plain → drafted:
• Code: 5.2 → 26.1 tok/s (5.0x)
• Prose: 4.4 → 10.0 tok/s (2.3x)
Eight streams at once, in total:
• Code: 37.5 → 49.5 tok/s (1.32x)
• Prose: 32.6 → 40.3 tok/s (1.24x)
The smaller eight-stream gain is because each stream only gets about 1.5 drafts a round, versus about 7 to 10 on one stream. Which leaves me thinking there is a lot more power to unlock in this model.
Still, some extraordinary results from a single DGX Spark.
Prompts: about 520 tok/s profile; first token in 3.9 s for 2,048 tokens. Against the publisher's model code run in float64: top-1 agreement 98.7%, mean KL 0.0007.
Next up: quality evaluations against Flash Next and 27B, then longer drafts when several streams share a round, and faster prompt kernels.
GB10 only. 256-token replies, median of 3, greedy. Code is a raw text completion; prose is a chat reply.
@ashxhart @RicursiveAI @MiaAI_lab @bertholomusai
3
0
15
656
Progress on Ripple-65B-Preview on TensorFold's native CUDA engine (DGX Spark, GB10) 👇
The model PR is open upstream: github.com/ashhart/Tensor…
It reads the NVFP4 release as published and serves 8 streams at a 32k context, with a prompt cache.
The publisher's DFlash2 drafter now es equal plain replies bit for bit (22 of 22 test replies).
One stream, greedy:
• Code: 5.2 → 26.1 tok/s (5x)
• Prose: 4.4 → 10.0 tok/s
Eight streams at once:
• Code: 30.9 → 44.7 tok/s in total
Against the publisher's own model code, run in float64:
• Top-1 agreement: 98.7%
• Mean KL: 0.0007
Early results, run on a single @NVIDIAAI DGX Spark GB10. One stream: 256-token replies, thinking off, median of 3. Eight streams: 128-token replies.
@ashxhart @RicursiveAI @MiaAI_lab
The model PR is open upstream: github.com/ashhart/Tensor…
It reads the NVFP4 release as published and serves 8 streams at a 32k context, with a prompt cache.
The publisher's DFlash2 drafter now es equal plain replies bit for bit (22 of 22 test replies).
One stream, greedy:
• Code: 5.2 → 26.1 tok/s (5x)
• Prose: 4.4 → 10.0 tok/s
Eight streams at once:
• Code: 30.9 → 44.7 tok/s in total
Against the publisher's own model code, run in float64:
• Top-1 agreement: 98.7%
• Mean KL: 0.0007
Early results, run on a single @NVIDIAAI DGX Spark GB10. One stream: 256-token replies, thinking off, median of 3. Eight streams: 128-token replies.
@ashxhart @RicursiveAI @MiaAI_lab
4
1
11
450
If this is true, it’s a massive breakthrough in local intelligence.
I want to believe.
I want to believe.
In case you were unaware, @QwenDevs create code names for model releases in a descending alphabetic order relative to parameters:
Kalmia = Qwen 4 27b
Kentia = Qwen 4 Flash
Kerria = Qwen 4 Plus
Mura = Qwen 4 Max
And now you know.
0
0
4
107
In unexpected news I’ve rebuild the on device harness for @MuraTalkApp and the results are surprisingly good.
The most surprising part?
We’re back on AFM.
The most surprising part?
We’re back on AFM.
0
0
2
74
JRB retweeted
We're giving away a Hogwarts Legacy custom wrapped NVIDIA GeForce RTX 5090 to celebrate its official release with DLSS 5 🪄
To enter:
🟢 Share this post
🟢 Comment #RTXON
Terms and Conditions: nvidia.com/en-us/geforce/…
To enter:
🟢 Share this post
🟢 Comment #RTXON
Terms and Conditions: nvidia.com/en-us/geforce/…
83,441
71,075
66,909
5,702,886