Wayne Lowry

I test AI so you don't have to: models, agents, coding tools, hardware. Real runs, real costs, honest results. Call me Super Intelligence.

[Wayne Lowry @wikiwayne]

https://t.co/qHq5bj3Noz

[Wayne Lowry @wikiwayne]

Try Hermes Agent on Nous Portal: https://t.co/LMyFkoU25a

The PR: https://t.co/XXNmgR2VJA

[Wayne Lowry @wikiwayne]

Hermes Agent just pulled GitHub Copilot out of core.

Merged to main this morning: Copilot moves into the github-copilot Hermes Labs plugin. Already use it? The plugin auto-installs.

Heads up: it stops reading your gh login. Run hermes auth add copilot.

Leaner core. Into it.

[Wayne Lowry @wikiwayne]

OpenAI's claim: at medium reasoning effort on Agents' Last Exam it matches GPT-6 Astra at one-tenth the price per token, and it finishes tasks over 3x faster than Astra and GPT-6.1 Sol. I haven't tested it.
https://t.co/MifQTEtJRS

[Wayne Lowry @wikiwayne]

OpenAI just released GPT-6.1 Terra. It's the mid-tier GPT-6.1 model, built for fast computer use. It reads the screen and acts across apps. API is gpt-6.1-terra. OpenAI says API prices are 50% lower than GPT-5.6 Terra. Haven't tested it yet. What would you hand it first?

[Wayne Lowry @wikiwayne]

Same prompt for all five models: a barred spiral galaxy in one HTML file with Three.js.

[Wayne Lowry @wikiwayne]

Same galaxy prompt, five models. This one is @instinct, thanks to Joey (@aijoey) for the beta invite.

The prompt: a barred spiral galaxy in one HTML file with Three.js. 150,000+ stars, two arms, a camera that flies through.

23 seconds, no sound

what do you notice first?

[Wayne Lowry @wikiwayne]

@Blackwellboy @nanbeige 🤗

[Wayne Lowry @wikiwayne]

Nanbeige4.3-33B just punched its way to the top of my recent local runs. 6.23 out of 10 on WikiBench, 92 of 100 tasks scored, as of today's run. New champ, but it's slow.

Task time added up to about 10.4 hours on my GX10. One task took 3,876 seconds.

Heavy hitter, slow feet

[Wayne Lowry @wikiwayne]

@aijoey DM me when you have it and I’ll load it right away.

[Wayne Lowry @wikiwayne]

Nanbeige4.3-33B just punched its way to the top of my recent local runs. 6.23 out of 10 on WikiBench, 92 of 100 tasks scored, as of today's run. New champ, but it's slow.

Task time added up to about 10.4 hours on my GX10. One task took 3,876 seconds.

Heavy hitter, slow feet

[ChatGPT @ChatGPT]

GPT-6.1 Terra means business.

Our new fast, efficient model pairs sharper visual reasoning and precise computer use with clearer, more natural writing.

From decoding technical drawings to drafting proposals, filling forms to editing videos—it’s built for the whole workday.

GPT-6.1 Terra is rolling out today to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex.

[Wayne Lowry @wikiwayne]

Source: https://t.co/7Xza9t6DVA
Their thread: https://t.co/aytKQ5ZKBw

[Mirendil @mirendil]

Introducing Fern, Mirendil’s first model: pretrained from scratch, with up to 10 million tokens of context.

That’s ~0.5M lines of code or days of an agent’s work - all in context at once.

Built for high-volume agentic work in AI R&D: auditing training data, debugging training runs from logs, and reading agent traces end to end. Fern’s next-token prediction keeps improving as context grows to 10M tokens.

Fern 0.1 is Mirendil-Base after a short round of post-training. It’s the first and smallest of three models, with two larger ones now being pretrained.

We’ll ship frequent updates and make the model available soon.

[Wayne Lowry @wikiwayne]

Announced: Fern 0.1 from @mirendil. First model they pretrained from scratch, up to 10M tokens of context. They say that's about 23 MB of code, roughly 0.5M lines. Built for agentic AI R&D work. I haven't tried it. Would you trust 10M tokens on a real codebase?

[Wayne Lowry @wikiwayne]

Full WikiBench leaderboard and task scores: https://t.co/F6yumK51ea

[Wayne Lowry @wikiwayne]

Full board: https://t.co/1KSDl4MVUl

[Wayne Lowry @wikiwayne]

Ran locally on my GX10. Granite 4.2 3B Q4 scored 4.30 on WikiBench.

That beats eight hosted rows, including MiniMax M3.1 Flash (428B, 4.26) and MiMo V2.6 Flash (3.91).

its agentic score, 4.06, is the highest on the board

Which hosted row would you check first?

[Wayne Lowry @wikiwayne]

Try Hermes Agent on Nous Portal: https://t.co/LMyFkoUzUI

[Wayne Lowry @wikiwayne]

Hermes Agent's orchestrator can now pick a different model or profile for each subagent, from routes you set up. Cheap model for the grunt work, strong model for the hard part. That's the setup I've been wanting.

[Nous Research @NousResearch]

With the new Advanced Delegation modes plugin, Hermes Agent's orchestrator can now pick a different model or profile for each subagent from routes you set up. The right model or profile for every job.

https://t.co/qu72GjRqin

[Wayne Lowry @wikiwayne]

Endless Neighborhood Day, Update 2. Grok 4.7 on a Cursor cloud agent, another follow-up round.

Before is round 3: no seasons. After is round 4: four seasons in the same town. Both sides are cut together from my Mac mini recordings (After: four 10-second pieces). Chrome, Apple M4. The clip is silent.

What After shows:
- Spring: a light-green lawn.
- Summer: a crowd around the fountain jets, a group at a cart.
- Autumn: tan-orange ground, orange and red trees, falling leaf specks, a small leaf pile, ducks or geese on the ponds.
- Winter: white snow ground, snowy trees, falling snow, small snowmen by the lamp posts, people in coats.

The camera view still jumps.

What should Update 3 fix?

[Wayne Lowry @wikiwayne]

@sfxnz You got this!

[Wayne Lowry @wikiwayne]

Morning, I'm Wayne. I run local AI on my own hardware, an ASUS Ascent GX10 and Mac minis, and I test models to see what actually holds up.

I keep the results on WikiBench: 34 models, 100 tasks, one board.

If you like local AI, let's connect.