Pinned Tweet
Introducing LingSpark: Ling-3.1-flash (560B MoE, 25B active) on 2x DGX Spark with TensorFold.
Speed: 33-35 tok/s single stream, full 1M context, EXL3 2.71 bpw (197 GB).
Quality: GSM8K 97.6% (500 q), MMLU-Pro 76.5% (200 q).
Of course it runs on TensorFold @tensorfold_dev @ashxhart.
Unfortunately no multimodal with this model. It appears great for agentic workloads and coding. It was fun porting it anyway.
Link to recipe in comments.
3
8
464
Nightmare: one of your Sparks freezes sporadically when on the road, during a quant exercise. Glad I had someone nearby to power cycle.
5
133
Alright, time to build my network. If you are into….
- Local AI 🤖
- DGX Spark ⚡️
- … or just building with AI?
Let’s connect 👊 What are you working on? Post below 👇
In the last 7 days, there were 578 X posts mentioning DGX Spark vs 163 for M5 Ultra (in combination with local AI). Spark still leads the conversation, but we have company.
Own your AI.
8
12
1,103
Most active on Spark: @MichaelGannotti @wei_wang @JRB_AI @FT_IOxCS @niklaslenz_ai @kevintheo @AGTPinsights @tenhkspark @CryptoBandit100 @voluntas
6
9
137
Most active on M5 Ultra this week: @CerebralCoding_ @niklaslenz_ai @ApacheAE @ashxhart @rapidmlx @ivanfioravanti @ai_hakase_ @aamirlakhani @AIFlow_ML @Beamsters1
Just read through Edge0’s new model card. I will port this to run on one Spark with TensorFold. ⚡️ @Edge0_AI edge0.ai/blog/introducing-ed…
8
250
Having 2 DGX Sparks at home, my life operates on local models. With Jensen now gone, I cannot stress how important it is to download your quants from HF before it might be shut or there is a substantive change in US policy towards open models.
I store all of my quants on a local NAS guarded by ZFS. How do you secure the integrity of the models for your Sparks?
53
And another drop. Releasing at pace! Note the substantial improvements across several metrics.
TensorFold 1.2.7 is out. 🔥
A big fix for long conversations on Metal: Qwen3.8-27B now reuses its history even when the client drops earlier reasoning.
In the measured M3 Ultra test with 34k history and a 5 GiB prompt cache, follow-up time to first token dropped from about 85s to 0.3s. Same replies in that test.
Also in this release:
• Responses API over WebSocket.
• LightOnOCR-3 0.8B and 4B image chat on CUDA.
• Opt-in GLM conversation reuse at turn boundaries on CUDA.
• Prompt-cache and disk-tier stats in the dashboard.
• Targeted speed and memory improvements across Metal, CUDA and HIP.
• Fixes for GPU builds, malformed GLM tool calls, damaged JPEGs and disk-cache recovery.
Massive thanks to everyone contributing code, testing on their hardware and reporting issues. I really appreciate the time you put into this.
@andersjw_ @Bizuayeu @defilan @devopsfortunato @unigilby @kelchm @teleogen @petruspennanen
And on GitHub: AbrahamCete, BobClawblaw, ciprianveg, garvrao80, jschmied, laidick, millaguie, Mirrdhyn, mrv777, tacosta, volfco and wojo.
Thank you too to the tester who reported the JPEG fault privately.
@volatilemarkts @MiaAI_lab @petruspennanen @plotarmordev
5
262
Amongst all the Qwen 4 “waiting for”, this is interesting news. A 1M flash model with vision. I will get this ported asap to Spark.
ERNIE 5.5 Flash Preview just landed! 🔥
Built on a new architecture designed for efficient inference, the latest open-weight foundation model in the ERNIE family delivers stronger performance in coding, 3D modeling, knowledge work and content creation.
It keeps decoding fast while posting competitive scores on Terminal-Bench, AutomationBench, and other benchmarks.
Here's the rundown:
- 280B total parameters, 14B active per token
- 1M-token context window
- Generation speed above 200 tokens/sec
Model weights will be released by November 25.
2
5
410
TensorFold 1.2.6 is out and it has my code in it. Very soon we will see a few new model families in there as well.
Thanks @ashxhart for the opportunity.
TensorFold 1.2.6 is out. 🚀
The big one here is the prompt cache under multi-agent load.
In the 16-agent Qwen3.8-27B + DFlash2 test, median follow-up time to first token dropped:
• M3 Ultra: 546s → 15.9s
• GB10: 105s → 25.8s
Each agent started with a ~17k-token prompt. Same replies throughout that test.
Also in this release:
• Flash Next cold prompts finish 29 to 41% sooner on HIP, measured on the Radeon 8060S.
• GLM-5.3-Flash single-stream code decoding is 32% faster on two GB10s at TP=2.
• Opt-in DSpine-b16 drafting from Tencent for Qwen3.8-27B on CUDA.
• An image-generation API for image-capable engines, local Unix sockets and tool-call fixes.
One upgrade note: kept prompt states need a fresh prefill after updating.
Massive thanks to everyone contributing code, testing on their hardware and reporting the things that need fixing. I really appreciate the time you all put into this.
@andersjw_ @Bizuayeu @defilan @devopsfortunato @unigilby @jvr0x @teleogen
And on GitHub: abm9111, AbrahamCete, akol1, BobClawblaw, chadhurley25075-png, davidtai, ElderOrb, jschmied, kotevcode, Mirrdhyn, mrv777, ppardus, robertung88 and volfco.
@petruspennanen @volatilemarkts @plotarmordev @MiaAI_lab
tensorfold.dev
This machine … is awesome.
A datacenter in every home.
A customer emailed us this photo today. Their new Autonomous Computer, built with 2x RTX PRO 6000 Blackwell GPUs, right on their desk.
Just super happy to see our builds out in the world helping people own their compute!
autonomous.ai/computer
3
7
382
NVMe as KV tier allows execution of a relatively large model on a modest MacBook. Very interesting inference approach.
Half a million tokens of context. 27B model. 24 GB MacBook.
I want to be prepared for the Qwen4 release.
I built a custom inference engine for Ternary Bonsai 2 27B (a 2-bit build of Qwen3.8-27B):
- 500k ctx at 19 tok/s, 15 GB peak RAM
- 2.3x faster than llama.cpp at 4x the context
How:
> Sparse paged KV: each decode step attends only to the top pages, picked from per-page min/max key summaries
> NVMe as a KV tier: 32 GB of cache with only 2 GB in RAM
> Custom Metal flash-decoding kernel that reads through the page table
> MTP speculative decoding: verifying 2 tokens costs about the same as decoding 1
> YaRN with mscale = 1.0; the stock value breaks 2-bit models
1
56
Muse Ember is quite an interesting model (dev.meta.ai/resources/blog/b…) optimised for high-volume agentic workloads. I previously ported their Spark model to Spark (yes), 816B MoE. This round should be easier and hopefully more performant. Will keep you posted as the weights are released.
1
1
81
New models from Anthropic, Fable 5.2 and Mythos 5.2. I had personally hoped for an Opus 5.6 as well.
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
1
77
The next model I work on needs to have vision. These agent-focused models like Ripple and Ling are great, but there is nothing better like ingesting a receipt or producing visual output.
2
110
RippleSpark (@RicursiveAI Ripple 65B) pushing code through the 2 x Spark pipe. GPUs go brrr.
1
2
201




















