Based on the benchmarks, it seems that Muse Glimmer barely edges out against Qwen3.6 27B, except for tool-calling skills (MCP, etc.). I wouldn't be surprised if they released it now because they are afraid they wouldn't beat Qwen3.8 27B.
I am glad they released it because I think we need a competitive culture of open weights that isn't just geopolitics.
But I have to say, I quite like the way Muse Glimmer thinks and talks. It's a cocky bastard in tone, but it's quite good, and its thinking traces are relatively terse.
Do AI companies make release plans based on upcoming other models like this? I would think all the processes that go into the repository and weight infrastructure pre-training, checkpointing, knowledge distillation, model compression, post training pipeline, ecosystem integrations, inference API, benchmarking, human eval/safety/alignment, docs, etc... all that dictates the release schedule.
Any company working in a competitive industry is generally aware of what their competitors are doing. PR is an important aspect to market success, so it factors into release schedule. It may not be the dominant factor given engineering constraints, but yea, it’s certainly a factor, and a large one at that.
There has been a long history of AI model releases made shortly before or after a major planned release by another company. Almost always to upstage or steal thunder.
Just recently, Minimax H3 released as open weights on the eve of Seedance 2.5 global availability. It's not as good, but it's good enough and it's completely open.
Flux 3, which is nowhere near as good as either, suddenly announced their release once news of these other two became public. They knew if they waited they'd be ignored. It didn't really help them much, unfortunately.
The LLM releases are even more rivalrous.
And don't forget all of the competing launches planned before Google IO or major release events.
Companies like to eat into the news and press cycle of their rivals.
BFL is in a rough spot here too. It’s pretty much looking like a repeat of the exact same situation they had when they released Flux2 at the same time Z Image Turbo came out and completely overshadowed their launch.
Minimax H3 can run exceptionally fast (10 minutes for a 15 second 0.5mp video and that's stock cuda 13), works on 16 GB VRAM GPUs, etc. If Flux3 is anything like Flux2, it’s going to require an absolute monster truck of a machine and still run significantly slower. Even if it’s a better model, that won’t matter as much if nobody releases any LoRAs or fine-tunes for it.
Not to mention BFL licensing often feels deceptively confusing and restrictive.
I've seen it here on HN (it's particularly noticeable via the /active page) multiple times. If Google, OpenAI or Anthropic release something significant, odds are good you'll see a headline from one of the others.
If you start counting since WaveNet or BERT, it's been ages. Especially when it feels like decades of advancements happen every single year, and rival labs are always trying to one up each other.
Yes, not every model release is reactionary to other labs. Either they had hints for the release of other models or they cut efforts in late stage testing of the models to hit these earlier release dates. There’s always some flexibility. And there’s certainly the incentive to cannibalize the news cycles for competitor models.
I could imagine pulling out all the stops to get a release over the finish line a week early if you're worried about being surpassed by another release
Yeah but you can probably have everything ready and then accelerate as necessary. Meta itself did this when releasing Llama 4, it was a really botched release right when they were feeling the heat from DeepSeek and others.
AI companies release models when they are ready - not on a set schedule. The steps are required to produce a release candidate, so a company can choose to bless the RC with the best metrics at any time - or alternatively continue (post/)training newer RCs if they feel they can do better, and they have time. When a model is ready for release is subjective, and can take into consideration what the competition is doing.
An obvious counter-example to schedules driving releases is the still unreleased Gemini 3.6 Pro.
the last few items there (benchmarking, human evaluation, docs) can be rushed or skipped by leadership if they want to beat comp. they probably spend a few weeks on those things normally
One window that can be shortened is working with software ecosystem and upstream partners; think day 0 on together, fireworks, Unsloth, etc. That obviously happens from partners getting embargoed weights early.
There is a finetune Qwen3.6-27B-Fable-Fus-711-UnHeretic-NM-DAU-NEO-MAX-NEO which seems to be as good at coding as vanilla Qwen, but way, way better at creative writing than Qwen and even better than Gemma 4 26 and 31b.
I saw that one in the "Popular models" sort at Hugging Face and tried it on some tasks I do frequently to compare models, and it feels damaged by the fine-tune, to me. It wrote security bugs into the code (probably just sloppy thinking, not intentional), it exhibited looping behavior in some configurations in llama.cpp, configurations I regularly use with the regular 27B, and it failed to write unit tests without being prompted (though the regular Qwen 27B tends to do so if it sees there are unit tests for everything in the repo). They have good benchmarks, but I'm not going to trust it. Also, that name is absolutely crazy.
> but way, way better at creative writing than Qwen and even better than Gemma 4 26 and 31b.
I suspect this is the only use-case I would consider...and I don't really have a use-case for "creative writing" that I would delegate to an LLM. I suppose for dialogue generation in games?
But yes, hard agree. Why on Earth would you ever want to write code with a model that is supposedly "jailbroken"? So it can put great backdoors into everything it touches? Pass.
I've noticed most fine-tunes, whether "heretic" models or something else, tend to be over-fitting, or something, at least some of the time, and get kind of chaotic. I want to believe normal folks with normal resources can be involved in this stuff, as I'm working on fine-tuned specialist models as we speak, but it seems like it takes notable investment and time. My first experiment was teaching a little Gemma 4 more to write more like me with a LoRA (like you, I don't want to use a model to write for me, but I did want training data that I could ethically use, and I've written several million words on the internet over the years), and it wasn't what I would call a success. It either wrote like an asshole (which I only do, like, 15% of the time) or it just borrowed a few of my quirks, like too many ellipses, if I applied it less heavily.
I use it to write short sci-fi stories. Life is not only about being an SDE.
> Why on Earth would you ever want to write code with a model that is supposedly "jailbroken"?
I need a model which decent enough at coding to write boilerplate for me, yet I do not need to unload in case I want to write an article or enguage in my writing hobby.Exactly why Gemma 4 are so popular.
> So it can put great backdoors into everything it touches?
> If we don't see something that's substantially better in the ~30B param space soon - it would appear we might've saturated that size with knowledge.
I wouldn't be quite so pessimistic. We may have saturated the current approach, but I think there's a lot still left in terms of compression, attention, active parameters, caching etc. etc.
I don’t think four months without a major breakthrough is cause to abandon all hope just yet. ;) The wild pace of LLM development is highly atypical, and we’re still in the ‘initial rush’ phase of development.
For contrast, the Newcomen steam engine (widely considered the first commercially useful engine) was used for over 60 years before the next major improvements. Now, 300 years later, we’re still finding ways to significantly improve heat engines.
> For contrast, the Newcomen steam engine (widely considered the first commercially useful engine) was used for over 60 years before the next major improvements. Now, 300 years later, we’re still finding ways to significantly improve heat engines.
Off-topic, but I stumbled upon the first Newcomen engine imported into Australia in a museum in Sydney and I was unexpectedly charmed (not an Engine Guy). It's large, but nothing like the awe of "mega-engineering", it's crude, but it clearly has such amazing utility (when compared to a reality without it) and it changed the world
I honestly expect that major advances in the open 30B dense space will take about a year, but expect incremental advances every couple of months from different developers in the meantime.
Qwen 3.6 27B was already a massive gift to smaller homelabs around the world; anything more is just a delightful surprise.
It will also be very interesting to see some direct head to head benchmarks between qwen 3.6 27B (let's say all at Q8 XK quantization, using the GGUF that unsloth publishes as a baseline) vs 3.8 27B. Particularly in tool use, terminal use.
The whole class of what can reasonably fit in a single GPU is an interesting category of LLM, and based on the results I've seen from 3.6 35B A3B and 27B versus what existed a year prior, it seems there's a lot of room for advancement.
I am working on a project where we have to classify customer calls into more than 10 categories. As the client wants everything locally I tried a few local LLMs. Gemma turned out to be the best model for this task. The classification accuracy is impressive, and the client is happy that I am using an American model.
Not necessarily. Except for SpaceXAI's (haha) models, there's less risk of reality distortion in Western models. Probably doesn't affect his use case but you wouldn't really want to ask Kimi K3 for a history of China for example.
The tokenizers are included in the open s̶o̶u̶r̶c̶e̶ weights releases; you wouldn’t be able to use the weights without the corresponding encoder/decoder, in fact.
I've been using Qwen3.6 35B A3B, and with reasoning turned on, I'd say 2/3 (give or take) of the tokens for a response are thinking tokens. Which at 70+ tps locally, that isn't that awful. I run an 80k context across 4-10 "agents" for my solo TTRPG, where Qwen is the GM, each NPC at a location, the director, and the narrator.
Each turn is about 45-60 seconds to generate all of the various responses. The GM and director have reasoning on, and the NPCs/Location/Narrator do not.
It's a fairly good "engine" for that. I'm not sure how a denser Qwen would do here regarding speed.
I like the tabletop RPG use case, and wanted to say: If your hardware likes it you should check out Gemma 4 for creative DMing use case. I found it to be much better at holding the plotlines and being creative on gaming turns. My experimental case was an audio-only Zork and Gemma 12B and even E4B were pretty good!
This sound very interesting, do you have any resource I could look at?
Me and my son did a very rudimentary (compared to yours) setup to play Paranoia, but this is at another level.
I'll go over my repo, and see if it is hiding any API keys and maybe make it public. The issue I have is it relies on a nuget package that also isn't live (its in my local nuget feed).
I'm not sure what all is needed to make that work for people.
I'm working on something similar. My biggest annoyance is that the overly-helpful LLM was making every die roll succeed. I ended up building some tooling around rolling dice. Also some tooling around character stats and inventory management, so those don't get lost in context compression.
It is on a single 3090, and that seems to be where it averages out. I'll get 85tps on turn 0, but then it settles down to low 70s within a few turns, but holds steady at that.
My issue currently is KV Cache, because I can't keep enough parallel caches running (4 is where I'm at), so TTFT (is that the initialism?) can be long when I have a particularly large scene (basically more than 2 NPCs).
But my harness does let me offload to any OpenAI compatible endpoint, I just prefer local cuz free.
Back in 2023 I started my own C# LLM library for doing tool calls and structured output, and over the years it has morphed bigger and bigger, and that is the backbone of almost all of my LLM-based projects.
I've never released it, but its easy to understand, and simple to add your own tools:
[AIDescription("Get current weather for a location")]
static string GetWeather(
[AIDescription("The city name")] string city,
[AIDescription("The country name")] string country,
[AIDescription("Temperature unit", ["C", "F"])] string unit = "C")
{
// make some API call to a weather API and return a string to the LLM
return $"The weather in {city}, {country} is 22°{unit} and sunny";
}
var chat = client.StartConversation("You are a helpful assistant with access to weather data.");
var response = await chat.SendAsync<string>("What's the weather in London?", GetWeather);
I'm sure plenty of better libraries exist for this now, but in 2023, I don't think any existed in the dotnet ecosystem. I've never released it though, because I've never "finished" it.
Not parent, but I use Goose for my non-handcrafted Qwen use cases, I’m also working on handcrafting as well. Goose was the only harness that didnt bloat context too much with system prompts (like openclaw) and I could get reasonable web search working with Qwen.
Just to play devil’s advocate: you can’t compare Qwen to a (proprietary/closed source) hosted model and deduce that Qwen is overthinking, as Qwen gives you the full reasoning/thinking trace while all the proprietary models now give you only a summary “to prevent distillation”, making it hard to properly compare apples to apples here.
People say Qwen overthinks because they analyzed the thinking traces, and Qwen finds the answer relatively quickly but then second guesses itself multiple times for another 20,000+ tokens. Regardless of what other models do, that's clearly overthinking.
This is most likely because the vast majority of the information the model absorbed during training was in Chinese. As a native Mandarin speaker, I frequently need to convert the prompt into English and output it in English in order to avoid that the model falls back into Chinese reasoning logic.
PS: Switching the thinking process from Chinese to English can also significantly circumvent certain self-censorship mechanisms built into the model.
You can use any message you want, but the model was tested to react reasonably well to the specific token sequence of "\nConsidering the limited time by the user, I have to give the solution based on the thinking directly now.\n</think>.\n\n" (from a Alibaba paper, struggling to find it now)
Since i have tools to prune context and run subagents, i just tell it to do either since both require summarization which is usually what it needs to avoid the long if...then chains
> Will be interesting to see how Qwen3.8 27B compares against this once it releases this week
Considering that Meta distills Qwen[1] (and should!), it'd be hilarious if Muse loses the head-to-head; the "distillation attack!!1!" people claimed distillation on release n-1 is enough to match the intelligence of the latest version.
Qwen3.6 is a definitive, significant downgrade from Qwen3.5 for creative writing and prose for example. Yes, it's better at agentic and coding, but it regresses in many non-coding areas compared to Qwen3.5.
Of course, I do expect the 3.8 ones to perform better for agentic coding.
One thing I would caution is staying out of the prediction market like this.
Tech tends to get boring when you judge current products against the hypothetical capabilities of unannounced products that may never ship. It's like comparing Nikon cameras against Canon camera rumours, or comparing iPhones against unannounced and therefore largely imaginary Samsungs.
- If they do a Qwen 3.8 35B A3B (and I hope they do because I love the 3.6 version)
- and if it beats 3.6 27B by all metrics
… then the local open weights world will be a better place.
But they have said nothing about it and they dropped several weight classes for 3.6, so who is to say they won't drop the 35B? And even if they don't, this is a tall order; why would the MoE tradeoffs no longer be apparent? (Again, I really like both the Qwen and Gemma MoEs)
FWIW I am enjoying testing Muse Glimmer — it's really quite impressive on chat, has nice terse and even amusing thinking traces, a bit of brass to it, and I'm hoping it will be good on agentic stuff.
Quantization awareness doesn’t change the size of the weights, just means it won’t degrade when quantized. QAT = quantization aware training. They will both be very similar in size at the same quant.
You're mixing up sizes of different quants. The 60GB is unquantized, and Qwen's unquantized size is around 54GB. Their sizes as like quantization levels are similar.
From my perspective it doesn't make sense to talk about the number of parameters. What matters is model size in bytes and its performance at that certain size.
Meta actually relesed official 4 bit quants in 17GB, but I haven't seen any indication that training was quant-aware, so the quants are not going to have same performance. 3.6 27B has official FP8 quant that AFAIR was trained with quantization awareness.
The best example is last year's gpt-oss which was released prequantized in mxfp4 so 20B parameter model was under 14GB and 120B was under 70GB right away.
That's exactly the point. We know short context knowledge stuff does not regress with quantization. But I expect agentic intelligence to suffer greatly.
If I were to pick one bench, I would like to compare quants on TerminalBench Hard. But then Glimmer already loses to 3.6 27B on it by a large margin.
EDIT: An open weight version of Muse Spark 1.2 is going to be released as well:
https://x.com/alexandr_wang/status/2086756152034066792
https://xcancel.com/alexandr_wang/status/2086756152034066792