OSS • Building local-first agent tooling | Harness optimization, local inference, small language models
Germany
- Tweets 346
- Following 469
- Followers 27
- Likes 1,051
My Gemini sub is finally paying off 🙏🙏🙏
Gemini 4 Argon is now live via the Gemini API and AI Studio for paid API and Google Ultra subscribers!
start building today: ai.studio/build
6
As mentioned in a previous tweet, I am currently using Astra and Gemini 3.8 Flash in Hermes in an advisor/executor setup.
I noticed that Gemini 3.8 Flash underperforms in coding tasks significantly when used as subagent, due to lacking context.
1
42
2/ But there are two PRs on the horizon who could solve the problem:
PR #106474 — forked subagent context
Snapshot the parent’s persisted conversation and pass it through conversation_history.
and...
1
20
Currently I am using Astra + Gemini 3.8 Flash for coding.
Astra as advisor and reviewer while Gemini acts as executor model.
It's wild how long one simple request takes, mostly Astra dealing with the hallucinations and sloppiness of Gemini.
1
20
Claude Opus 5.5 and GPT-6 Sol & Luna launched yesterday.
Benchmark results paint a clear picture now:
Opus 5.5 generally leads GPT-6 Sol and espc Luna on independent intelligence, coding and knowledge-work benchmarks, while Sol & Luna win on cost-efficiency and automation tasks
1
1
41
Pierre Martinow retweeted
The main appeal of open-source agent harnesses isn't that they are free, but that we can inspect what they are doing on our computers.
124
80
908
39,685
Grok 4.7 and MiMo-V2.6-Pro came out at the same time and both landing on the same score of the AA-index.
16-18th rank, both scoring 46.
MiMo-V2.6-Pro secured the top spot for open weights models now.
Looking forward to more benchmark results TBA over the next days.
52
wow – what's going on with Gemini?
Round-trip latency ~45-50s, 1-4 t/s.
Gemini 3.8 Flash turned to Gemini 3.8 Slow.
2
5
421
MUST SEE: Amazing moment as President Trump calls Nvidia CEO Jensen Huang while he's on stage at the All-In Summit.
@POTUS on AI Doomerism: “I'm telling you, it's all a hoax… and we're not going to let that happen.”
21
Anthropic provided pretty damning evidence that Chinese labs use US models to train their own.
If US labs actually slow down for safety now, does that stall China too? Hard to distill models that haven't been built yet.
13
Open weights flash models become really interesting for local inference now.
You could run them on old 2nd or 3rd gen AMD EPYC rigs and get DGX spark level memory bandwidth.
1P/2P, 8-16 memory channels, DDR4 3200, 200-400 GB/s.
INT4 quantizations for CPU inference.
2
99
Running an agent harness, memory system and models locally is no joke.
That's a full-time job, my friend!
43
Thank you Google, I was genuinely waiting for this 🙏
Gemini 3.8 Flash (High) by @GoogleDeepMind has reshaped the Pareto frontier for Agent Arena!
This model is priced at $0.75/$3.75 per MToken (input/output).
In Agent Arena, Gemini 3.8 Flash (High) has a median cost of $0.22 per task and +5.94% net improvement. This performance is on par with Grok 4.5 by @spacexai and GLM 5.2 (Max) by @Zai_org, but at a 44-50% lower price point.
- Gemini 3.8 Flash (High) (+5.94% at $0.22/task)
- Grok 4.5 (+6.17% at $0.39/task)
- GLM 5.2 (Max) (+6.23% at $0.44/task)
The release of Gemini 3.8 Flash (High) has landed Gemini on the Agent Arena Pareto for the first time.
Congrats again to the team @GoogleDeepMind!
99