Building the neovim of agent harnesses. bone-agent. 26, swe in finance.
New York, USA
Joined April 2026
- Tweets3K
- Following241
- Followers186
- Likes2K
Im having grok file a complaint with chatgpt about codex. Its gonna argue with the other ai for me. Im so excited.
What is with the dated model version numbers
Codex at capacity last night. At capacity this morning. I cant get a single message through. Thw audacity to reset my limits during all this....
Codex is still have capacity issues. Haven't had that happen a single time with claude, or deepseek, or grok...
everyones acting like this changes the game and isn't another grok bot dot. It's gonna do the same things. I'd be happy to be wrong.
My feel benchmark is preferring q4 27b over iq3 qwen flash. Flash just drones on and on calling bash never finishing a task, 27b gets to the point. Idk if its a quant thing. I can run them at almost the same speed, but 4x more concurrent sessions of 27b.
I tried running a vowel bench of qwen flash iq3 and 27b q4 to see which is best on my hardware. Neither could get it right on a single attempt no harness. Is that normal?
do you think deepseek engineers use Opus 5.5 for complex tasks? DSV4.1 is good but its a flash model in the end of the day
I get 140tps on qwen flash iq3 on 2x 3090s, but my first test ran 43 shell calls in a row and couldn't solve the problem before sol 6.1 in the same harness. I really want to like it but I think the quant is too small
Maybe with the release of deepseek infrastra some of these other labs can get their tps somewhere respectable?
Didn't think it was possible but I can run qwen 3.8 flash on 2x 3090s with only 32gb ddr5. There's hope for the ddr5 poors.
Finishing up my personal android app, both ssh client with a gui and tui for my custom agent harness.
I still find myself using the tui. It just feels more crisp
X is doing some magic with their mobile app network and caching I can scroll for 20 min with comments on barely any connection.
I used deepseek v4.1 flash for 90% of my work at work today just to get a feel for it. I have unlimited usage of all models. Very impressed, and only spent 5 dollars. Share holder value 📈📈📈
I learned i love spawning swarms of deepseek subagents. 20 all at once to brute force test a new feature. Get basic qa done in like 60 seconds.
Higher output but the cache rate being 0.015 is gonna make this super cheap. Now to measure how to efficient it actually is. Not hard to beat dsv4.1 here
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
