building backend infra for ai apps • senior tech lead • i actually ship stuff
Joined January 2016
- Tweets 2,757
- Following 441
- Followers 1,558
- Likes 102,486
152 Photos and videos
Jun 12
Using for a few hours, will share how it feels compared to k2.6
Jun 12
🌘 Kimi-K2.7-Code, our latest coding model, is now released and open-sourced!
🔷 Improved coding & agent performance over K2.6: 21.8% on Kimi Code Bench v2, 11.0% on Program Bench, and 31.5% on MLS Bench Lite.
🔷 Reasoning efficiency: Less overthinking, with 30% lower reasoning-token usage compared to K2.6.
🔷 Long-horizon coding: Improved instruction following, higher end-to-end coding task success rates.
⚡️ 6x High-Speed Mode coming soon!
🔌 Available today via Kimi API and Kimi Code.
🔗 Kimi Code: kimi.com/code
🔗 API: platform.moonshot.ai
16
Jun 5
Now that fire pass is sunsetting, is there any similar offering with no token limit? maybe without weekly limit?
27
Jun 5
Sad to see such offerings go away. I’ve been enjoying this, no token meter, endless generation but i can understand it’s not sustainable. Will wait for V3 @FireworksAI_HQ
31
May 29
Let’s go! I’ve been rooting for stepfun. It’s a great model and could be greater in future. Kudos to the team
May 29
⚡️ Step 3.7 Flash is here: The new frontier is agent efficiency.
#1 ClawEval-1.1 (67.1), #1 SimpleVQA Search (79.2), #2 SWE-PRO (56.3), 95.3 on V* Python. Open weights under Apache 2.0.
Built for agentic, coding, search, and multimodal workflows — balancing speed, cost, and reliable execution.
- 400 TPS. 198B sparse MoE, ~11B active. 256K context, 3 reasoning levels.
- Understands UIs, charts, docs, images — then writes code or calls tools to act on what it sees.
- Web visual search reaches further: more sources, deeper follow-up.
- Reliable tool use — less drift, fewer broken toolcalls. 98% on τ²-bench across all difficulty levels.
- Works with Claude Code, KiloCode, Hermes Agent, OpenClaw, and protocols like MCP.
- Runs locally on Mac Studio M4 Max, DGX Spark, AMD AI Max 395.
GitHub: github.com/stepfun-ai/Step-3…
HuggingFace: huggingface.co/stepfun-ai/St…
GGUF: huggingface.co/stepfun-ai/St…
ModelScope: modelscope.cn/models/stepfun…
API: platform.stepfun.ai
Blog: static.stepfun.com/blog/step…
31
May 28
I don't run Claude or Codex. No subscription, no tab open, no need at all. Kimi 2.6 Turbo via Firepass and Opencode Go handle everything I throw at them, vibe coding, production bugs, work tasks. More tokens than I can burn. What are you using Claude / Codex for?
92
May 28
Qwen3.7-max is really really good.
Prompt: /frontend-design implement a todo app which looks like car dashboard in /tmp
Result:todo-car-dash.surge.sh/
1
39
May 25
Kimi 2.6 has been my daily and exclusive model for quite sometime. I don’t really miss opus or codex.
May 23
> be Kimi Founder
> 32 years old
> peers are choosing corporate jobs
> you're building AI infrastructure from scratch
> China. no English press. no Western hype.
> raise $2B. hit $20B valuation. quietly.
> nobody outside China writes your name
> invent a new optimizer. 2x more efficient than the industry standard.
> build 300-agent parallel systems. 4,000 steps. 12 hours straight.
> open source. free. beats the models everyone pays $200/month for.
> one afternoon you sit down
> record 40 minutes
> give the entire playbook away for free
> the math. the architecture. the decisions.
> a week later Western developers find it
> "wait. this exists?"
> "wait. it's free?"
> "wait. it beats Claude?"
> you were already on the next version
> they were just catching up
> different game.
1
100
May 23
Just tried qwopus3.6-35b-a3b-v1 on my MBP 32GB RAM and its surprisingly good. Not fast by any means but its something I can run in the background, fire off a one-off task, and leave. Perfect for quick code reviews and small refactors while I work on other things.
1
71
May 23
The setup is dead simple. Download it through LM Studio, load the model, and start chatting. No API keys, no rate limits, no usage anxiety. Just your machine and the model. For solo devs working on side projects this changes the entire cost equation.
1
55
May 23
A year ago local models at this size were basically unusable. Now I have a 35B parameter model running quietly on my laptop handling real tasks. The local model community has made insane progress and its only accelerating. The future is definitely not just cloud APIs.
21
May 23
Problem: every AI coding tool writes the same bland code. Claude Code, Cursor, Codex — they all produce "correct" code that looks nothing like how I actually write. Different naming, different patterns, different structure. The agent doesnt know my taste.
4
3
187
May 23
The end goal: AI that writes like you, not like everyone. Every developer has a style. Mine is in 146 repos. Now the agent can learn it. Building in public at github.com/shahidcodes/neuro…
3
43
May 23
Deepseek v4 pro is gonna change pricing game for everyone. It is a model with immense cap for 25x cheaper than anything similar
23
May 23
DeepSeek going permanent 75% off! Just migrated from $200/mo API costs to $12. The model is commodity, the value is what you build on top of it.
May 22
We are making our discount permanent! 🎉
Enjoy building with DeepSeek-V4-Pro and bring your innovative ideas to life! 🚀
57