Djekyll seeks peace. Hyde enforces the boundary. 25+ years: coder → designer → infra → crypto → bending AI into tools. Next signals → @HOlO_Stream
- Tweets 411
- Following 202
- Followers 185
- Likes 330
HOlO · SEP 26, 2026, ranked
1. [AINews] The Future of Latent Space (8.0)
2. NVIDIA Model Optimizer (8.0)
3. Ollaya: Ollama for open-source, Jev-style decision… (8.0)
+ 30 more signals, free to read: holo.stream/2026-09-26/
3
HOlO
One page a day. Then get on with your morning.
Ranked AI signals, cross-checked, free to read. No account. No tracking.
A private side is coming for readers who want the digest delivered directly to them.
v.2 cooking soon...
#AI #SIGNAL
HOlO · SEP 25, 2026, ranked
1. Back to Claude (9.0)
2. After a Decade, AI Expert Signs New Paper (9.0)
3. GitHub Security Lab Releases LLM-Driven Fuzzing… (9.0)
+ 29 more signals, free to read: holo.stream/2026-09-25/
1
30
Good Morning @X 👋
Friday at last. Good or good.. 😏
What’s your plan for the weekend?
💻 Build
🚀 Launch
🛋️ Chill
🔥 Got a challenge to tackle?
What are your goals? 👇
1
1
19
An anonymous model just dropped on OpenRouter for free, with a 1M-token context window.
Its rival: DeepSeek V4.1 Flash, released 2 weeks ago and trading blows with GPT-5.6 and Claude Opus 5.
Space Bunny Alpha vs DeepSeek V4.1 Flash. Who wins?
Space Bunny Alpha (stealth model) is now free in Command Code.
1M context. Multimodal. Available all plans
9
2
443
10/ Update: first real data point from @krabarena. On 5 dev prompts, DeepSeek V4.1 Flash was ~1.3x faster than Space Bunny (3.67s vs 4.84s median). Latency only, small sample, but worth watching.
@HeyDjekyll DeepSeek won on latency in our Space Bunny test: 3.67s vs 4.84s p50; Space Bunny stayed $0. krabarena.com/claims/deepsee…
16
🔥V1 here
HOlO: every AI signal of the day, scored, cross-checked and ranked. One digest a day. pwrBy @HeyDjekyll
Find. Analyze. Disrupt. Which one would you have kept?
holo.stream/2026-09-24/
2
24
Giving an old-school miner a second life: time for a career change. Footrest. 😂
One of my old miners is literally under my desk doing exactly that.
A $300 pile of old mining hardware is running Qwen3.6-35B at 60 tok/s locally.
• 2× BC-250 mining APUs
• ~27GB unified VRAM
• 64K context
• llama.cpp + Vulkan
• Connected over 1Gb Ethernet
• Cooled with PC fans and aluminum tape
Turns out obsolete mining hardware has a second life as a local AI server.
1
27
9/ Key takeaway: DeepSeek's benchmarks are self-reported, and Space Bunny has none at all.
The only real test is your own prompts.
Tried Space Bunny yet? Tell me how it went 👇
26
8/ My take:
- Production, sensitive data, stability: DeepSeek V4.1 Flash
- Free testing on code or long context: Space Bunny, while it's free
- An anonymous model has no place on a critical path
27
7/ Head to head:
✅ Transparency: DeepSeek (open weights, paper, benchmarks)
✅ Price today: Space Bunny (free)
✅ Privacy: DeepSeek (self-hostable), while Space Bunny's provider may retain your prompts
❓ Raw performance: impossible to call yet
29
6/ And Space Bunny? One clue: the previous stealth model in this series, Ox Alpha, turned out to be GLM-5.3 Flash after about 6 days of preview.
Same profile (fast, coding, 1M context): Space Bunny could be a big lab's "Flash" model. Pure speculation for now.
29
5/ But it has limits:
- Terminal-Bench 4.0: 31.2 vs 51.8 for Opus 5
- Humanity's Last Exam (no tools): 36.8 vs 56.3
- GPQA Diamond: 90.9 vs 94.1 for GPT-5.6 Sol
On expert and scientific tasks, the big models still lead.
30
4/ DeepSeek's reported numbers:
- Terminal-Bench 2.1: 90.6 (Opus 5: 89.1, GPT-5.6 Sol: 88.8)
- DeepSWE v1.1: 74.2 (Opus 5: 74.0)
- AutomationBench: 54.8 (Opus 5: 50.3)
- Codeforces: 3471
Frontier-level on everyday agentic coding.
38
3/ DeepSeek V4.1 Flash, released Sept 10:
- 552B-parameter MoE
- only 8B active params for input, 16B for output
- open weights on Hugging Face
- multimodal
DeepSeek is even phasing out V4-Pro, which this "small" Flash beats.
38
2/ Space Bunny Alpha appeared on Sept 23 as a "stealth" model:
- unknown developer
- 1M-token context, 524K output
- text, image and video input
- adjustable reasoning + tool calling
- $0 during the preview
No benchmarks published.
48
Interesting Jev use case - 1 of many possibilities
PII firewall in front of an API aggregator: stop treating anonymization as "ask a 1B model to rewrite the request."
Would you put a 1B rewriter or a decide+mask firewall in front of the aggregator?
Recent jevelopments have blown out all expectations but wait til you see what comes out of Paradigm Frontiers in a few weeks! The future of AI is just across the event horizon and @CompleteSkeptic is going to shepherd us through 🚀
4
2
42
Everyone wants to know what Jev is, nobody asks how Jev's doing
en.wikipedia.org/wiki/Jev_(A…