ฅʕ •ﻌ•ʔฅ

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@stevibe thx but i m on Mac, so vllm wont work.
Would be cool if we can see models playing minesweeper with it!

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@stevibe oh i mean which inferencer?
saw u mention it requires a HTTP server exposing /predict-upload

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@r1cA18 おすすめ

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@djnmrjnvc Is it M3 Max?
Q8 i got like 6 t/s in plain
so i tried to bench on q4

[ฅʕ •ﻌ•ʔฅ @AllenH__]

🤔 llama-server (4f13cb7)
M3 Max 128G
-m unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-UD-Q4_K_XL.gguf \
-ngl all -c 8192 -ub 2048 -np 1 -fa on -ctk q8_0 -ctv q8_0 \
--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0 \
--reasoning off \

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@RohitP_tweets Yes all these with spec type mtp
spec draft n max 6 at the last row

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@ivanfioravanti not seeing much improvements on M3 Max 128G
https://t.co/oszqdIk7tz

[ฅʕ •ﻌ•ʔฅ @AllenH__]

🤔 llama-server (4f13cb7)
M3 Max 128G
-m unsloth/Qwen3.6-27B-MTP-GGUF/Qwen3.6-27B-UD-Q4_K_XL.gguf \
-ngl all -c 8192 -ub 2048 -np 1 -fa on -ctk q8_0 -ctv q8_0 \
--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0 \
--reasoning off \

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@lucasmeijer Was using pi + local Qwen3.6 35B + customized CDP script (pi wrote) - works alright.

then i realize it has official cli too - even smoother
https://t.co/zOU4nQTKvz

[stevibe @stevibe]

Six open-source LLMs. One sliding puzzle. A brutal test of long-horizon reasoning and tool calling.

Five of them broke. One didn't.

I gave each model a move_tile tool and a scrambled 3×3 board, then asked it to solve the puzzle through pure turn-by-turn reasoning. The deeper the scramble, the more brutal the search.

Five runs per depth, best run kept. A model fails the round if it exceeds 6x the optimal move count.

> Depth 5:
Everyone solves it. Yawn.

> Depth 10:
GLM 5.1 melts down. 43 moves. Cut.

> Depth 12:
Gemma4 26B loses the plot, shuffling tiles in circles. Gone.

> Depth 15:
The wall. DeepSeek V4 Flash, out. DeepSeek V4 Pro, out. Gemma4, out again. GLM 5.1, out.

Two survivors: Qwen3.6 35B-A3B, and Kimi K2.6 with an 11-move solve that looked like cheating.

> Depth 18:
Same two. Everyone else hallucinating tiles that weren't there.

> Depth 22:
Final boss. Kimi, flawless for five rounds, finally cracks. 81 moves. Still scrambled. DeepSeek V4 Pro limps home at 90.

Qwen3.6 35B-A3B solves it in 36.

The smallest model in the room. ~3B active params. Fits on a single 3090. It beat everything.

Kimi was elegant. Qwen3.6 was unstoppable.

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@stevibe 27B too!
https://t.co/CbPkkT3GqS

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@Youssofal_ @ivanfioravanti @AlexJonesax Any plan to merge your work into these mainstream inferencers?

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@Youssofal_ @devindolar Tested but doesnt seem support streaming response in pi? and somehow TPS is lower than builtin chat.

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@Prince_Canuma @LambdaAPI M3 Max 128GB
+ tps: pp 18 ; gen 31 tps
~ peak mem usage 97GB
- require special branch of both mlx_lm and hf transformers
- loop in thinking in long prompts (>200 tokens)

suggest: wait for WIP PRs to be merged

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@jun_song everytime when a new model launch have to type these differently:
`-ctk q8_0 -ctv q8_0 -b 4096 -ngl all -fa on -n 2
--temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.0 `

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@kis are these from 4 bit models?

[ฅʕ •ﻌ•ʔฅ @AllenH__]

@betup_ @0xSero still no idea, back to pi and all good

[Qwen @Alibaba_Qwen]

🚀 Meet Qwen3.6-27B, our latest dense, open-source model, packing flagship-level coding power!

Yes, 27B, and Qwen3.6-27B punches way above its weight. 👇

What's new:
🧠 Outstanding agentic coding — surpasses Qwen3.5-397B-A17B across all major coding benchmarks
💡 Strong reasoning across text & multimodal tasks
🔄 Supports thinking & non-thinking modes
✅ Apache 2.0 — fully open, fully yours

Smaller model. Bigger results. Community's favorite. ❤️
We can't wait to see what you build with Qwen3.6-27B! 👀

🔗👇
Blog: https://t.co/P2Zx7FwMxB
Qwen Studio: https://t.co/c4vm4LuZrU
Github: https://t.co/zKDEbv0R4U
Hugging Face:
https://t.co/N67hyzxvfr
https://t.co/SSdtbWRDap
ModelScope:
https://t.co/xODf1pj9kw
https://t.co/xXhoqlJ2AB