We are here to ensure our partners become the signal of innovation in the noise of the technology markets.
- Tweets 460
- Following 107
- Followers 1,310
- Likes 615
Certainly agree, and glad we are building pinnacle.signal65.com to address specifically these questions!
Nvidia is leaning hard into the trend of enterprise companies wanting to run their own open models in their own data centers and also trying to make on-device, private AI agents a thing.
spr.ly/6016BHr816
1
3
213
From server-level performance to rack-level impact.
Independent testing from @Signal_65 highlights how Arm AGI CPU performs across key agentic AI workloads, demonstrating the importance of CPU performance as AI infrastructure scales.
Take a closer look: okt.to/LPaNF3
2
4
33
2,748
Signal65 retweeted
We have the first tests for the $ARM AGI CPU and the results are quite positive.
In June I wrote that the @Arm AGI CPU still had no independent benchmarks. The story is the rack, not just the chip: 1.99x the modeled work per rack of @AMD's EPYC 9755, built on a 1.20x per-chip edge.
@Signal_65 published the study across seven CPU workloads agents lean on, from MySQL to SWE-bench.
-Per chip, AGI CPU wins five of seven, losing MySQL and Milvus.
-Per rack, the multiplier is power. Arm rates a two-socket AGI CPU server at 950 to 1,050 watts versus 1,700 for EPYC 9755, so 21 to 23 fit in a 24 kW rack against 13.
-SWE-bench goes from 1.13x per chip to 2.00x per rack.
That backs my call since June 2025 that agentic inference puts the CPU and performance per watt back at the center.
Footnotes:
-Arm set the methodology, published last week at @OpenComputePrj Global Summit, and ran early systems in its own facilities. Signal65 watched the runs live, parsed the raw files and rebuilt the math.
-Both EPYC parts are Turin, launched two years ago this month, not Venice, which AMD said will roll out starting in Q4.
Next I want AGI CPU versus Venice, rack versus rack.
Enterprises scaling agentic AI can get about twice the CPU work out of the same rack power with the @Arm AGI CPU, based on @Signal_65 modeled rack analysis.
Full report: signal65.com/research/ai/arm…
That matters because power and floor space now cap how far AI infrastructure can grow. We compared the AGI CPU against the leading x86 processor across seven agentic and enterprise workloads.
Key findings from our analysis:
➡️ 1.99x average modeled performance per 24 kW rack, ranging from 1.43x to 3.19x by workload
➡️ 21 to 23 AGI CPU servers fit in a rack, against 13 x86 servers, based on rated server power of 950 to 1,050 W vs. 1,700 W
➡️ Higher per-chip performance on five of seven workloads, averaging 1.20x
➡️ About half the racks for the same work on web serving, agent replay and model-serving overhead
7
9
59
7,148
🎇 NVFP4 version of Muse Spark Open supports 5x more agents 🎇
On one 8x NVIDIA B300 node the NVFP4 build with DSpark speculative decoding holds 112 concurrent agents, where the BF16 build holds 24, in our testing.
That is 4.7x the agents from the same B300 node and the same model, and the 4-bit NVFP4 weights give up under 3% of the PINNACLE Model Score for it, with retrieval fidelity on the 128K corpus identical at 97.2%.
A correct multi-step enterprise task costs $0.12 on the NVFP4 build against $0.37 on BF16. Both builds are live on the results views at pinnacle.signal65.com
@nvidia @AIatMeta
2
6
44,722
Enterprises that are weighing Meta’s open-weight Muse Spark Open against the hosted Muse Spark 1.3 API have two numbers that frame the decision.
The open weights finish a correct multi-step enterprise task for 7.1x less than the hosted model, even though the hosted model earns a higher Signal65 PINNACLE Model Score.
➡️ $0.12 per correct task for Muse Spark Open in NVFP4 on a rented 8× NVIDIA B300 node, against $0.88 for the hosted Muse Spark 1.3 at Meta’s list rates, in our testing
➡️ The hosted model makes 55% fewer weighted errors. The gap sits on the frontier-tier agentic jobs, 95.0% finished against 79.2%
➡️ Muse Wonder 1.0, Meta’s hosted flagship, costs $1.70 per correct task, 13.8x the open weights, for a Model Score in the top ten
An enterprise that needs open weights it can host and audit gets a correct task for a seventh of the API price but gives up the hardest tier of agentic work to do it.
Both Muse Spark Open rows are live on the cost view at pinnacle.signal65.com
2
318
Signal65 retweeted
Meta published a detailed engineering look at Prometheus, the gigawatt-scale supercluster in Ohio that trained its @AIatMeta Muse models and is scaling toward roughly 1 million GPUs (!!) running as a single machine across buildings linked over distances of more than 100 km.
According to the story, joint work with @nvidia cut training interruptions 99% from the mid-2025 peak, with each NVL72 rack running a 64-GPU slice and holding eight GPUs as in-rack spares. Every GPU platform in the cluster is NVIDIA-based, from GB200 and GB300 today with Vera Rubin NVL72 coming up next, and that of co-engineering keeps NVIDIA firmly in the lead for frontier training.
The AMD position at @Meta, which spans millions of EPYC CPUs and a 6 GW Instinct agreement focused on inference, is unaffected, simply because we see demand for compute still outrunning what any single supplier can deliver.
engineering.fb.com/2026/10/1…
4
21
1,410
The @NVIDIARTXSpark brings 128GB of unified memory and a Blackwell RTX GPU to Windows laptops, making it practical to run large agent models locally without per-token cloud costs.
Signal65 President @ryanshrout walks through how NVIDIA positions the N1X.
➡️ Up to 128GB of LPDDR5X unified memory, which NVIDIA says supports models up to 200B parameters
➡️ A claimed 4.6x lead over the Intel Core Ultra X9 388H on the Artificial Analysis AgentPerf-local benchmark
➡️ About 2x the performance of an RTX 5070 laptop at the same 60W, per NVIDIA
➡️ A token-economics case of $20 a month in electricity versus $1,200 in API fees, based on an always-on desktop
Signal65 testing of the @Surface Laptop Ultra is underway, with independent results to follow.
3
214
💡 MUSE SPARK OPEN BECOMES CLEAR BEST US-BASED OPEN WEIGHTS MODEL 💡
Lots of enterprises want a frontier-class open model they can run on their own hardware from a US company. @AIatMeta Muse Spark Open, the open-weight release of Muse Spark, is the best US open-weight model on Signal65 PINNACLE, by a wide margin, and fourth among open models worldwide. (pinnacle.signal65.com)
What were our key takeaways in the intelligence scoring?
➡️ PINNACLE Model Score 3,367, #30 of 82 scored builds open and hosted
➡️ #1 of 16 US open-weight models. It makes 57% fewer weighted errors than the next US open model, Meta's own Muse-Glimmer-30B, and 69% fewer than Nemotron-3 Ultra, the best from outside Meta
➡️ #4 of 35 open-weight models worldwide. The three above it are Alibaba's Qwen3.8-2.4T-A95B, Qwen3.8-Flash-Next and DeepSeek-V4.1-Flash; it sits ahead of GLM-5.2, DeepSeek-V4-Flash-0731 and Kimi-K3
➡️ It completes our agentic tasks in 20 median rounds per job against 32 for the hosted Muse Spark 1.3 model, with 14% fewer output tokens per correct answer (meaning it is more efficient in practice)
It finishes 90.7% of the multi-step enterprise jobs, and no other US open model on the board finishes more than 72.5%. For an enterprise that needs open weights it can host and audit, this is the US model to start from.
4
10
26,311
Enterprises planning agent deployments need to size the CPU as carefully as the GPU. Between model responses, agents run shell commands, compile code, execute tests, and manage containers, and that execution sets how many agents a server can sustain and what each task costs.
Signal65 tested AWS m8i (Intel Xeon 6), m7i (4th Gen Intel Xeon Scalable), and m8a (5th Gen AMD EPYC) instances on Terminal Bench 2, using recorded inference responses to remove LLM variance and isolate the CPU-side work.
Key findings from our testing:
➡️ m8i completed tests 6.9% to 11.0% faster than m7i at 2.2% to 6.6% lower compute cost
➡️ m8i test cost was at parity or better versus m8a at four of five concurrency levels, with a 5.5% advantage at 20 concurrent tasks
➡️ With a multi-process harness, average CPU utilization on m8i.16xlarge climbed to roughly 54%, up from under 10% with standard single-process runs
Full report: signal65.com/research/ai/ben… @intel
2
1
6
83,378
NVIDIA launched Nemotron 3.5 this morning, and the first independent agentic results for the family are on the Signal65 PINNACLE leaderboard now. For an enterprise weighing an open-weight model for agent work, the question is whether the new 120B-class model finishes the job and what it costs to get there.
What we found in our testing:
➡️ Nemotron 3.5 Super makes 43% fewer weighted errors than the Nemotron 3 Super it replaces, at the same model size
➡️ It finishes 4x the share of multi-step enterprise jobs, in 7.5x fewer rounds, with 123x fewer billable tokens per correct task
➡️ It reaches 90% of the Nemotron 3 Ultra 550B score on a tenth of the tokens, with better retrieval fidelity
➡️ It still lands 55th of 80 builds, and three of our seven enterprise tasks account for nearly all of the gap
Results views: pinnacle.signal65.com @NVIDIAAI
13
42,497
Signal65 retweeted
$GOOGL comes out on top. Best performing and not most expensive model for agentic tasks. 👏🏻
🚨🚨 Google is back at the top of the agentic PINNACLE leaderboard! 🚨🚨
Gemini 4 Argon leads Signal65 PINNACLE (pinnacle.signal65.com) at its default effort, and at Google's standard $4/$20 rate a correct task costs $0.87, below the $1.08 and $1.51 of the two models it passed.
➡️ 13% fewer weighted errors than Claude Sonnet 5.5 at max effort and 35% fewer than GPT-6 Astra at max in our testing, with 100% of the multi-step enterprise jobs finished and 98.9% on the 128K retrieval corpus
➡️ 14 median rounds per job and 10,549 output tokens per correct answer, against 29 and 63,020 for Sonnet 5.5 max, which is why a job costs less on Argon at twice Sonnet's input rate
➡️ The top of the board has moved up 53% since September 5th and the price of a correct task on it is 43% lower
We priced @GoogleAI Argon at the standard rate on purpose. Google's $2/$10 promotion runs through November 30th and would read 43 cents; the number that maintains through December is the one on the board.
2
5
90
6,977
🚨🚨 Google is back at the top of the agentic PINNACLE leaderboard! 🚨🚨
Gemini 4 Argon leads Signal65 PINNACLE (pinnacle.signal65.com) at its default effort, and at Google's standard $4/$20 rate a correct task costs $0.87, below the $1.08 and $1.51 of the two models it passed.
➡️ 13% fewer weighted errors than Claude Sonnet 5.5 at max effort and 35% fewer than GPT-6 Astra at max in our testing, with 100% of the multi-step enterprise jobs finished and 98.9% on the 128K retrieval corpus
➡️ 14 median rounds per job and 10,549 output tokens per correct answer, against 29 and 63,020 for Sonnet 5.5 max, which is why a job costs less on Argon at twice Sonnet's input rate
➡️ The top of the board has moved up 53% since September 5th and the price of a correct task on it is 43% lower
We priced @GoogleAI Argon at the standard rate on purpose. Google's $2/$10 promotion runs through November 30th and would read 43 cents; the number that maintains through December is the one on the board.
1
24
39,167
📰Meta has a flagship agent model again. 📰
Muse Wonder 1.0 enters the top ten of the Signal65 PINNACLE board (pinnacle.signal65.com) above Claude Opus 5.5 at both efforts, GPT-6 Sol max and Grok 4.6 in our testing, a 1.5x step over Muse Spark 1.3, and the one thing holding it off the cost frontier is the price - something @Meta can change.
➡️ 33% fewer weighted errors than Spark, in 19 median rounds against 32, with 49% fewer output tokens per correct answer
➡️ 99.6% of the multi-step enterprise jobs finished, 97.6% on the 128K retrieval corpus, fabrication 0.0%, level with the best on the board
➡️ $1.70 per correct task at Meta's $4/$20 list rate, 1.9x Spark and 5.6x GPT-6.1 Sol max, which scores higher and holds the frontier step above it
Intelligence is the hard part and Meta seems to have it; the rate card is the adjustable part. At Spark's $1.25/$4.25 card the same work would cost 49 cents and sit on the frontier.
Every score and every price of a correct task is live at pinnacle.signal65.com
2
6
35,901
Signal65 retweeted
Find more information on AI Field Day #AIFD9 here: techfieldday.com/event/aifd9…
@TechFieldDay @TheFuturumGroup @TechstrongTV @SFoskett @Signal_65 #AIFD9 #AI
1
3
83
Signal65 retweeted
AI Field Day 9 #AIFD9 continues now! 📅 October 14 - 15, 2026!
Tune in to @Signal_65 🔴 LIVE at 10:00 AM US/Pacific Time on YouTube, LinkedIn, Techstrong TV, or our website.
🔺 YouTube: piped.video/techfieldday
🔹 LinkedIn: linkedin.com/company/tech-fi…
@TechFieldDay @TheFuturumGroup @TechstrongTV @SFoskett #AIFD9 #AI
1
1
2
231
Enterprises scaling agentic AI can get about twice the CPU work out of the same rack power with the @Arm AGI CPU, based on @Signal_65 modeled rack analysis.
Full report: signal65.com/research/ai/arm…
That matters because power and floor space now cap how far AI infrastructure can grow. We compared the AGI CPU against the leading x86 processor across seven agentic and enterprise workloads.
Key findings from our analysis:
➡️ 1.99x average modeled performance per 24 kW rack, ranging from 1.43x to 3.19x by workload
➡️ 21 to 23 AGI CPU servers fit in a rack, against 13 x86 servers, based on rated server power of 950 to 1,050 W vs. 1,700 W
➡️ Higher per-chip performance on five of seven workloads, averaging 1.20x
➡️ About half the racks for the same work on web serving, agent replay and model-serving overhead
2
3
20
75,194
Arm AGI CPU benchmarks are here!
Independent testing from @Signal_65 shows strong performance across agentic AI workloads, translating server-level results into rack-level outcomes.
Take a closer look at the results: okt.to/rUzkM4
2
20
92
7,142
Signal65 retweeted
A model small enough to run on a 128GB desktop like the @NVIDIARTXSpark now does better agentic work than Claude Opus 5 in our @Signal_65 PINNACLE testing, and it lands within 5% of a 2.4-trillion-parameter flagship that took 64 GB300 GPUs to serve!
That changes the math on devices like RTX Spark and DGX Spark. The hosted frontier still finishes tasks faster and still makes 2x to 3x fewer errors at max effort. For agents that can run in the background, though, local is now a serious option on cost and data control. Both the frontier and local models are improving every release, which makes hybrid designs look better each time.
This week @Microsoft and @NVIDIA put local AI on a pedastal with new Windows updates and the official release of @NVIDIARTXSpark systems.
Whether this works at all and is USEFUL depends on how much correct agentic work a model small enough to run locally can deliver, and Signal65 PINNACLE measures to the improvement in those smaller models where this could do that.
➡️ Qwen3.8-Flash-Next scores slightly BETTER than Claude Opus 5 (released in July) and GPT-6 Sol (September) at medium effort on PINNACLE agentic work
➡️ It lands within 5% of the 2.4-trillion-parameter Qwen3.8 flagship with about 13x fewer parameters
➡️ A 4-bit build fits in 128GB, the unified memory of a DGX Spark and the top RTX Spark configurations. We have not scored it yet, but other 4-bit and 8-bit builds have landed close to their 16-bit versions in our testing, so we expect at most a small drop
These scores measure quality, not speed. Qwen3.8-Flash-Next needs about 2.7x the output tokens of Opus 5 per task, and a hosted model like Opus 5 will certainly finish agentic work faster than Flash-Next on a laptop or desktop.
Local AI trades that time for no per-token bill and data that stays on the device, and for many agent jobs it will be "ast enough." PINNACLE also scores agentic enterprise work rather than general knowledge or writing, where larger models can still hold an edge. pinnacle.signal65.com
2
1
4
828
This week @Microsoft and @NVIDIA put local AI on a pedastal with new Windows updates and the official release of @NVIDIARTXSpark systems.
Whether this works at all and is USEFUL depends on how much correct agentic work a model small enough to run locally can deliver, and Signal65 PINNACLE measures to the improvement in those smaller models where this could do that.
➡️ Qwen3.8-Flash-Next scores slightly BETTER than Claude Opus 5 (released in July) and GPT-6 Sol (September) at medium effort on PINNACLE agentic work
➡️ It lands within 5% of the 2.4-trillion-parameter Qwen3.8 flagship with about 13x fewer parameters
➡️ A 4-bit build fits in 128GB, the unified memory of a DGX Spark and the top RTX Spark configurations. We have not scored it yet, but other 4-bit and 8-bit builds have landed close to their 16-bit versions in our testing, so we expect at most a small drop
These scores measure quality, not speed. Qwen3.8-Flash-Next needs about 2.7x the output tokens of Opus 5 per task, and a hosted model like Opus 5 will certainly finish agentic work faster than Flash-Next on a laptop or desktop.
Local AI trades that time for no per-token bill and data that stays on the device, and for many agent jobs it will be "ast enough." PINNACLE also scores agentic enterprise work rather than general knowledge or writing, where larger models can still hold an edge. pinnacle.signal65.com
4
23,005