Can we expect the Qwen3.6-35B MOE version of this?
Thank you for all your great work. I’ve tested this on my M1 Max, and it’s a bit slow. So, I’m wondering if there’s a chance of getting this in the Qwen3.6-35B MOE model?
Yes, in fact I tried to download it, and it keeps timing out on the last tensor. Will keep working on it, the Tess models are excellent, and work great in a merge.
Yes, in fact I tried to download it, and it keeps timing out on the last tensor. Will keep working on it, the Tess models are excellent, and work great in a merge.
Wait does that mean there's an issue with this specific qx64-hi-mlx quant of the 27b? or just download issues?
or are you referring to a 35b-a3b finetune variant.
sorry, i'm eager to test some of these on my m5 max after running the DavidAU 27b on my 5090.
What I meant, the 35B MoE from Migel Tissera is stalling at download 5.21G/6.06G on the last tensor.
There is a MoE in this range, but doesn't have the lore and Tess. Once I get to the source, I will add Tess to this one.
https://huggingface.co/nightmedia/Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-Coder-1M-qx86-hi-mlx
I believe I’ve tested the one you mentioned above, but I’ve settled with this one: https://huggingface.co/nightmedia/Qwen3.6-35B-A3B-Fable-Holo3.1-Text-1M-qx64-hi-mlx as my daily driver. It performs better in my setup, using less memory and exerting less pressure.
What I’ve noticed is that the Coder version works better in Python, not in other programming languages. I’m happy to be wrong about this information.
No, that is perfectly reasonable considering the traces in the MoE are usually Python code, being the language less likely to fail when mined in one prompt.
The 9B series has custom made, backend engineering code lessons, in Rust, Haskell, Golang, and of course Python.
Those are built with a completely different set of questions than the usual, and more aimed on learning good programming practices, not coding itself. With Tess added, I would hope it do well on multi-step coding. Same thing goes for the Bradbury-Orwell, it's less related to what's in the book, more on reasoning and wording, so it sounds less like a Qwen.
https://huggingface.co/nightmedia/Qwen3.5-9B-Holodeck-Tess-Bradbury-Orwell-q8-hi-mlx
Thank you for the explanation. Since the Qwen3.5 model is older, will it generate code using the latest version? For instance, if I used to code for the .NET backend, will it use .NET 10 and C# 14, or will it use .NET 6 or 8?
I expect all traces from Qwopus, Fable, and Holo3.1(especially that one) to be hitting on the most recent frameworks when they distill from Cloud.
That is, of course, if there are .NET traces in there. Also, having models in the merge that are 3.6 based helps with what the Qwen knows about languages, so it will be a mix of experiences, so to speak. You make a good point that training with sources that have an early cutoff date, some stuff will be too stack overflow-ish :)
When I trained the 9B with custom traces, I used a 3.6 based MoE as a teacher, mainly because I wanted to preserve the Qwen language in the distills. When I was doing trace analysis with Gemini, all feedback came with glowing reviews--and yes, Gemini does get very nitpicky sometimes, but on those traces she had nothing to complain about.
What I meant, the 35B MoE from Migel Tissera is stalling at download
5.21G/6.06Gon the last tensor.There is a MoE in this range, but doesn't have the lore and Tess. Once I get to the source, I will add Tess to this one.
https://huggingface.co/nightmedia/Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-Coder-1M-qx86-hi-mlx
Ahhh, okay gotcha. I'll give this one a go then, thank you.
I know it's a little unrelated to this specific quant, but i wanted to test against this https://github.com/basecompute/baseRT which supposedly has better mac performance for pp/tg, but only works for BF16. So i put it against nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 since Tess's BF16 is gated, and technically it built a .base file but it blew up on tool calls. i threw the froggeric chat template in and did a fresh convert on it and it didnt blow up on the tool calls, but it didnt actually work on them either, which explains issue #21 on their repo.
so, long story short, i'll still try this tess quant via rapid-mlx expecting poor performance on the m5 max ~600GB/sec unified memory, but will attempt the 35B hoping we'll get a super heretic/fable/davidAU special that you guys made magical in DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF 😃
Thank you for the explanation. Since the Qwen3.5 model is older, will it generate code using the latest version? For instance, if I used to code for the .NET backend, will it use .NET 10 and C# 14, or will it use .NET 6 or 8?
honestly in my experience if you set the technical requirements and coding details in your project's AGENTS.md with strict instructions for your model to use websearch or to pull methods/details from the latest code details (sorry i dont know proper .net terminology, i only do python/rust programming + frontend work), then that will be enough.
my local models pull fresh rust crates and will look at source code changes, will go out to pypi, will pull other repos for requirements, will go out to pull docs on the latest vite, etc, just fine and it's able to code using the latest terminologies, and i just have searxng/camofox/huggingface mcps.
You might try de-ROPEing the model back to 256K, and it will be faster. I made it 1M just because some people asked, but you're rarely going to use it in a production setting, and it will be better for it.
Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-Coder
qx64-hi 0.658,0.833,0.895,0.779,0.432,0.823,0.734
qx86-hi 0.644,0.836,0.897,0.780,0.446,0.818,0.736
A bit counter-intuitive, but on reasoning tasks, the smaller qx64-hi has better arc numbers. In some cases that matters
Here are some examples from the 9B TNG training set:
Rust: Lazy Futures Retrieval: Provide a pattern that spins up three async tasks and binds their future results using lazy stream evaluation combinators.
Rust: Worker Load Balancing: Implement an array processing batcher that splits a list of items evenly across an arbitrary number of active async workers.
Rust: Concurrent Resource Pools: Create an allocation ring that leases shared resources safely across asynchronous threads using an asynchronous semaphore.
Rust: Thread Latency Spikes: Write an instrumentation wrapper that tracks individual async completion latencies and logs variances over a strict ceiling.
Worf is coding in Rust, and his prime directive is that Zero-Copy is the way, and memory is sacred. On watch are Garak, Quark, and Q that chime with smartass comments to keep Worf on his toes. This happens both in the think tag and in the output. All coding lessons are with antagonists in the room, to create the tension necessary, so there will be a joke at the end about how silly that challenge was :)
The PKD questions are to root both PKD lore and typical tasks where he would be an expert, so when you ask for Philip to help design one of these, he's done it before.
PKD: The Clockwork Pre-Crime Flash: Create a deterministic clock sync routine that keeps the pre-cognitive buffers aligned with the real-time simulation tick.
PKD: The Can-D Layout: Design a multi-user synchronization layer where a group of users ingests a simulated compound to share a singular, highly vivid layout environment.
PKD: The Chew-Z Infinite Vector: Architect a state machine for an unpredictable, un-bounded simulation space where the user's local thoughts instantly rewrite the physical terrain rules.
PKD: The JJ-180 Paranoia Overlay: Create a rendering engine modification that maps a layer of deep suspicion and hidden corporate conspiracies over every neutral environmental asset.
PKD: The Down-er State Throttle: Provide an internal system loop that mirrors the neurological slowing of a user under heavy down-ers, matching the physics tick to their delayed cognition.
You might try de-ROPEing the model back to 256K, and it will be faster. I made it 1M just because some people asked, but you're rarely going to use it in a production setting, and it will be better for it.
Qwen3.6-35B-A3B-Fable-Holo3.1-Qwopus-Coder qx64-hi 0.658,0.833,0.895,0.779,0.432,0.823,0.734 qx86-hi 0.644,0.836,0.897,0.780,0.446,0.818,0.736A bit counter-intuitive, but on reasoning tasks, the smaller qx64-hi has better arc numbers. In some cases that matters
fyi can i ask how to de-ROPE the model from the 1m down to 256k context?
In the config, find these:
"max_position_embeddings": 1048576,
...
"factor": 4.0,
"original_max_position_embeddings": 262144,
"rope_type": "yarn"
replace with:
"max_position_embeddings": 262144,
...
"rope_type": "default"
The tokenizer_config.json also has a line:
"model_max_length": 1048576,
that you can change to
"model_max_length": 262144,