invented plan mode • leading @usehamster, the ai native workspace for product teams • creator @taskmasterai
task-master.dev
Joined June 2011
- Tweets11.7K
- Following1.9K
- Followers18.3K
- Likes127K
Pinned Tweet
Sick of @cursor_ai rewriting good code or going in circles?
Introducing Task Master ✨
A CLI that turns your PRD into a local task management system for Cursor Agent
Graduate from building cute little apps to ambitious projects without overwhelming yourself or Cursor
👀👇
Eyal Toledano retweeted
Parts of the SDLC have become a bottleneck. At @usehamster we're on a mission to break down these walls.
Rate of change in a codebase has gone crazy. It's hard to review. It doesn't have to be.
If you like the idea of using Glass but don’t want to wait, try @usehamster
It’s the multiplayer self-improving, self-healing workspace we’ve been working on for more than a year that streamlines the full SDLC
And it’s S-tier
- full direction loop for the whole team
- discovery and alignment across living artifacts
- cloud delivery across briefs and initiatives
- shared method and skills + a curated library of methods and skills to one click install
- routines for automating work across the team
- self-writing blueprints describing product that auto update as your team ships
- self-healing context graph with team decision intelligence and intent lineage to answer what was built, when, why and by who and who made the decisions
Built with love by the team who built @TaskmasterAI
Eyal Toledano retweeted
horde.sh/ built this a few weeks ago, it’s working great for us internally
New @usehamster release brings native PR review and tours + @github routines for pre- and post- PR automations
Speed!!! 🐹
New @usehamster update 🐹
Native PR reviews with your team! No jumping between Hamster and GitHub. Read diffs, leave comments + approve or request changes from Hamster directly
+ Trigger routines from @github activity. Easily handle PR follow-ups with zero manual intervention
More Qwen3.8-Flash-Next breakthroughs, releasing soon in a new tiered model and supported by a custom MLX engine built by the @HamsterResearch team
- you only ever download the bf16 tiered model (upcoming)
- pMLX can quantize it up/down bf16/q8/q4/q3 on the fly
- 130-150 tok/s aggregate decode (crucial for sub agents)
- adjustable n-gram streaming from nvme
- adjustable experts residency ratio
- prune/REAP experts on the fly
- powered by pMLX
hundreds of config possibilities self-tuned from your hardware
examples for code-editing (which this decodes 1.5x faster):
- q4 @ 32k context, 95% experts in ram, n-gram in memory, 87 tok/s burst, 58 tok/s sustained (72gb memory), 149 tok/s aggregate
- q4 @ 256k context, 95% experts in ram, n-gram in memory, 87 tok/s burst, 50 tok/s sustained (74gb memory), 130 tok/s aggregate
- q3 @ 32k context, 95% experts in ram, n-gram in memory, 111 tok/s burst, 74 tok/s sustained (54gb memory), 190 tok/s aggregate
lowest memory footprint config: 12gb 🤯
- q3 @ 8k context, 30% experts in ram, n-gram on disk, 16 tok/s burst, 10 tok/s sustained (12gb memory), 29 tok/s aggregate
launch targets achieved, now packaging it up for you!
Another Qwen3.8-Flash-Next breakthrough
I stored 60% of experts on disk which stream to memory on-demand. Similar technique as n-gram streaming
Just eval'd with FULL experts (no prune) running on just 37gb of memory and decoding at 40 tok/s on my M4 Max
Hold on to your socks!
"which quant should I download?" is a question you may never have to answer again
the team @HamsterResearch has figured out how to kill it with pMLX. download once at full precision (bf16) and the engine re-fits it to your machine on the fly, based on the job you give it
tell it two things: how much context you need, and the slowest speed you'll accept. it reads your mac and picks the quantization plus how many experts stay in ram vs. stream from disk. no need to download a smaller quant or figure out which quant fits your hardware. same model for every use case and the config changes based on what you need
I ran this on my own M4 Max and Qwen3.8-Flash-Next splits into two configs.
short context, under 32k: full bf16, ~85% of experts in ram, fast at full precision since it fits my ram
long context, 64k to 256k: keep ~70% bf16 and stream the rest from SSD at 22 tok/s, or drop to q8 and get 40 tok/s. I pick per task and the model itself never changes.
there are many possibilities since I have the ram to spar
- if i need speed, Q4 100% resident (74gb ram @ 62 tok/s)
- if i need balance, Q8 95% resident (76gb @ 38 tok/s)
- if i need accuracy, bf16 70% resident (96gb @ 16 tok/s)
given whatever RAM you have (16/32/64/128/256 GB) + your context + your min speed, the engine picks the precision (bf16→q8→q4) and the expert-residency/paging split that fits your box and maximizes quality & speed
last thing to optimize is speed. there are so many things we want to power with open models at @usehamster and these 180b-300b class models have the potential to play a big role in that
"which quant should I download?" is a question you may never have to answer again
the team @HamsterResearch has figured out how to kill it with pMLX. download once at full precision (bf16) and the engine re-fits it to your machine on the fly, based on the job you give it
tell it two things: how much context you need, and the slowest speed you'll accept. it reads your mac and picks the quantization plus how many experts stay in ram vs. stream from disk. no need to download a smaller quant or figure out which quant fits your hardware. same model for every use case and the config changes based on what you need
I ran this on my own M4 Max and Qwen3.8-Flash-Next splits into two configs.
short context, under 32k: full bf16, ~85% of experts in ram, fast at full precision since it fits my ram
long context, 64k to 256k: keep ~70% bf16 and stream the rest from SSD at 22 tok/s, or drop to q8 and get 40 tok/s. I pick per task and the model itself never changes.
there are many possibilities since I have the ram to spar
- if i need speed, Q4 100% resident (74gb ram @ 62 tok/s)
- if i need balance, Q8 95% resident (76gb @ 38 tok/s)
- if i need accuracy, bf16 70% resident (96gb @ 16 tok/s)
given whatever RAM you have (16/32/64/128/256 GB) + your context + your min speed, the engine picks the precision (bf16→q8→q4) and the expert-residency/paging split that fits your box and maximizes quality & speed
last thing to optimize is speed. there are so many things we want to power with open models at @usehamster and these 180b-300b class models have the potential to play a big role in that
Now available:
BF16 & GGUF for Qwen3.8-Flash-Next-REAP-288
GGUF (Q4_K_M 78GB / Q5_K_M 87GB / Q8_0 116GB)
huggingface.co/sh0wie/Qwen3.…
bf16 (231GB)
huggingface.co/sh0wie/Qwen3.…
MXFP4 and NVFP4 uploading as well (last REAP releases)
Tiered model + pMLX release comes next
Enjoy!
From the @HamsterResearch lab:
Introducing Qwen3.8-Flash-Next-REAP-288-MLX-4bit, a 180B-class model running on just 39gb of memory
- MLX-native 4-bit 60% smaller than stock q4
- Pruned 512→288 experts via REAP
- 91.5% HumanEval (vs 93.9% stock)
@huggingface links below ↓
+ 2 variants:
Qwen3.8-Flash-Next-REAP-288-MLX-8bit
- 8 bit, 90.9% HumanEval, 70 GB resident
- q8 27% smaller than the q4 stock model
huggingface.co/sh0wie/Qwen3.…
Qwen3.8-Flash-Next-REAP-384-MLX-8bit
- 4 bit with 384 experts, 92.1% HumanEval, 51 GB resident
huggingface.co/sh0wie/Qwen3.…
This model could go fast. REALLY FAST.
MLX inference today (mlx_lm, mlx_vlm and other friends) is constrained by components it depends on. That's why I'm writing a pure MLX inference engine - to get the most out of our machines.
It opens up techniques and possibilities (i.e. kernel fusion) that would otherwise be impossible to develop.
So far, two big wins in sight with the new engine:
- Reached 65 tok/s with zero accuracy loss, aiming for 100+
- Expert paging brings the model down from 39gb to 20gb!
We're exploring much of this at @usehamster and will house most of this research and future open source releases from the @hamsterresearch account, so make sure to follow it.
Send this to a RAM-constrained friend. Help is coming.
Qwen3.8-Flash-Next now fits on a 48gb macbook air
Got q4 running on MLX using only 39gb of memory and decoding at 28 tok/s 🔥
Also q8 running on just 50gb of memory
I stacked two tricks so far:
1. REAP pruning cut the expert pool from 68gb to 35gb
2. the 51b n-gram table now lives on the ssd instead of ram
My patch memmaps the shards and reads ~100 bytes per lookup straight off disk, dequantized on cpu and verified bit-exact against the gpu path
Qwen3.8-Flash-Next at 288 experts with its PLE table on nvme loads in 6 seconds, 39gb resident, 28 tok/s decode, 600 tok/s prefill. Logits identical to the ram-resident version in every way
97gb stock Q4
down to 68gb if pruned to 288 experts
down to 39gb if pruned to 288 experts with the table on disk
humaneval scoring 91.5% with 288 experts vs. 93.9% for stock
I'm not done. next:
1. MTP drafter for speculative decode
2. n-gram lookup drafter
both are built, benchmarking them now
now this is interesting for the MLX side because this technique could stack with expert pruning i just showed works
mmap + page-in-touched-rows is already how mmap works on a mac with unified memory. there's no PCIe staging in between (thats part of the point)
it might be worth trying to move that process off metal to the cpu which would shave another 25gb off vram
the double-dip:
- REAP shrinks the 63gb expert pool
- offloaded PLE removes the 25gb table from residency and onto the SSD
stock model size goes down to 72gb with PLE offload (which is what the GGUF technique is for this model)
BUT the REAP-256 pruned model could land at 40gb! that would mean a 180b class model on 48gb of vram 🔥
It runs. Qwen3.8-Flash-Next on a single DGX Spark, day one.
Unsloth Q4_K_XL, llama.cpp (unmerged PR), full 262k context. The 51B n-gram table stays on NVMe via mmap, only touched rows page in.
~22 tok/s decode, ~500-660 tok/s prefill. And the number I was worried about: barely 1-3 page faults per token, dropping as context grows. The SSD trick costs almost nothing.
Decode falls only 13% from 226 to 19k prompt tokens. Long context is nearly free on this architecture.
Numbers in the screenshot. Next: making it faster without touching quality.
built the qwen3.8-flash-next pruning curve
it tells you the tradeoff between size and accuracy. every point is a pruned model scored against the unpruned one on held-out prompts
256 experts is the sweet spot: half the size before it buckles
a few interesting tidbits:
- i tried cutting random experts at the same size to check and the model fell apart. picking which experts survive matters hugely
- i also tried the q2 quant with full experts (63gb) and the output was pure garbage. the q4 quant with half the experts is miles ahead
