Model Requests & Suggestions
Suggest a model for a future handcrafted MiniPlus or NanoPlus release.
Please include:
- Upstream model link
- Intended use or strengths
- Why it would be valuable at a MiniPlus or NanoPlus footprint
I will review requests based on demand, architecture, local-inference practicality, and whether the model would benefit from a custom tensor-by-tensor quantization.
CyberTiel
- Link: peculiar-ragdoll/Cyber-Tiel-Coder-35B-A3B-GGUF-MTP (https://huggingface.co/peculiar-ragdoll/Cyber-Tiel-Coder-35B-A3B-GGUF-MTP)
- Just like Tiel , agentic coding tasks, but stronger, less thinking(in my opinion) and uncensored
- Miniplus like what you did on Tiel
The MiniPlus V2.1 and NanoPlus quantizations are already underway; they will be available within an hour at mostโfollow me to get notified.
Wow
Thanks for your hard works and very fast reply, really impressive, this is the fastest model request i've seen on hf (or maybe you had already worked on this model from before :))
I will definitely check and test for if the Apex Miniplus CyberTiel work good too
Request for next model.
Reason: I've personally used k2 and it's the closest thing to match or exceed Qwen intelligence:
IFM/K2-Horizon-MoVA-36B-A4B
I personally tried the K2-Horizon, the smaller versions and that MoVA, and they didn't convince me because of the cost of the KV Cache.
I could barely get it to 8K context in Q4_0, not enough to properly test it, is it really any good?
For me, the fact that it had such an absurdly expensive KV cache was enough to completely rule it out. Did you guys test it in real-world use?
You were using Q4_0 KV?
How much VRAM on the rig?
I assume your models were tested on 24gb ram?
I tried it on an RTX 3090 and I simply didn't find that series of models appealing; the KV cache was too expensive.
Gotcha, no worries. Thanks for the responses, love your work.
Gotcha, no worries. Thanks for the responses, love your work.
If you find another model that you consider good, don't hesitate to let me know so I can quote it immediately.
are you interested in some thing called "Ternary-Bonsai-2-27B-gguf" (*edit:oh i tried result really bad, rather choose ornith 1.5)
Yeah, despite Prism-ML claiming over 90% retention, in practice they perform terribly. Ornith 1.5 is just way better.
among all 35b model, which provide you the best performance in agentic coding tasks ?
atm i'm using Tiel/CyberTiel
Tiel and CyberTiel are definitely great for that, but Occamy fits right in there as well.
While Tiel is very focused on direct syntax and fast tool execution, Occamy handles agentic coding equally well while offering a bit more balance on broader architectural context, complex debugging, and multi-file reasoning.
If you want to give it a spin to compare against your Tiel setup, here is the MiniPlus V2.1 release:
https://huggingface.co/IsValorum/Occamy-1.0-APEX-I-MiniPlus-V2.1-GGUF
Between Tiel/CyberTiel and Occamy, you're pretty much looking at the best options in the 35B tier for agentic workflows right now.
Yeah i see
But maybe i still go with tiel until qwen 4 35b a3b(just guess lol)
The author optimize this model kv cache really good so it take less vram/ram for large context amount
And it have mtp too
After apex miniplus exist so i change to mini plus when need quality, and comeback to iq3 xxs when need 262k context
Totally fair point. That DeltaNet KV cache efficiency plus MTP makes it a great daily driver for long context. Enjoy Tiel!
If you're dropping to flat IQ3_XXS just to fit the full 262K context, check out the NanoPlus release of Tiel:
https://huggingface.co/IsValorum/Tiel-Coder-35B-A3B-APEX-I-NanoPlus-GGUF
That's actually why I built NanoPlus: to give you that lower memory footprint for massive context while still maintaining Q4_K_M/L-class practical quality instead of the flat degradation of standard IQ3_XXS, keeping routers in F32 and output in Q6_K.
Yeah, but i think miniplus with 131k still work good, even can upto 180-190k if i reduce some prefill boost parameter and disable mtp
Thoughts on the new European Model? Aleph-Alpha/Kolibri-1 https://huggingface.co/Aleph-Alpha/Kolibri-1
78B with ~4B active, gets decent benchmark
Hi @userr99 ,
I looked into it as soon as it dropped today, but unfortunately it is physically impossible to make GGUFs for it right now.
The main blocker is that Aleph Alpha built a completely custom architecture from scratch (Kolibri1ForCausalLM). llama.cpp has zero support for it: the GGUF converter will simply crash because it cannot map the tensors, and the C++ engine has no kernels to execute its specific 4:1 sliding-window attention (4 SWA layers for every 1 GQA layer) or route across 384 micro-experts per layer (that is 19,200 expert FFN blocks across its 50 layers). It does not even run on standard vLLM without installing Aleph Alpha's own proprietary plugin package (aleph-alpha-inference).
On top of the architecture gap, the base weights are 145.5 GB in BF16. Even if someone in the community spends weeks writing a custom C++ PR to get kolibri1 into llama.cpp, a 78B MoE would still land around 32 to 35 GB in a MiniPlus quant, meaning it would not fit into a single 24 GB VRAM card anyway without spilling into RAM.
So for now, until upstream llama.cpp actually writes and merges full native support for the architecture, it is completely out of reach for GGUF.
i see something called "Xing4.0-29B-A4B" ,does it have any advantage when comparing to 35b models (coding, speed, weight,...)
| Benchmark / Metric | Xing4.0-29B-A4B | Tiel-Coder-35B-A3B (Ornith-1.5) | Qwen3.6-35B-A3B (Base Reference) | Direct Comparison |
|---|---|---|---|---|
| Terminal-Bench 2.1 (Terminus-2) | 57.50 | 67.80 | 52.50 | Tiel +10.3 points ahead in interactive bash terminal execution |
| Terminal-Bench 2.1 (Claude Code) | (Not reported) | 68.50 | 49.20 | Tiel specialized for CLI coding agent harnesses |
| SWE-bench Verified (GitHub issue solves) | 75.00% | 79.00% | 73.40% | Tiel +4.0% higher resolution rate on real production codebases |
| SWE-bench Multilingual (Polyglot coding) | 66.00% | 71.40% | 67.20% | Tiel +5.4% stronger multi-language support |
| SWE-bench Pro (Professional repo issues) | (Not reported) | 59.60% | 49.50% | Tiel +10.1% over stock Qwen3.6 |
| DeepSWE (Complex architectural tasks) | (Not reported) | 22.00% | 0.00% | Tiel solves 22% where baseline 35Bs score 0% |
| SWE-bench-Live (25 live problems, Q4 GGUF) | (Not tested) | 12 / 25 (48.0%) | 8 / 25 (32.0%) | Tiel matches Claude Opus 4.6 medium (8.6 min median solve time) |
| Claw-Eval (Agentic workflows) | 76.55 | 67.20 | 74.54 | Xing4.0 leads in multi-step planning and tool-calling flows |
| MiniPlus V2.1 Model Weight | 13.76 GB (3.42 BPW) | 14.75 GB (3.45 BPW) | 14.75 GB | Xing4.0 saves approx. 1 GB VRAM on base model weights |
| NanoPlus Model Weight | 11.74 GB (2.92 BPW) | 12.60 GB (2.95 BPW) | 12.60 GB | Xing4.0 saves approx. 0.86 GB VRAM on ultra-compact weights |
| Attention Architecture | MLA + mHC (Latent compression) | GQA (Grouped-Query) | GQA | Xing4.0 compresses KV representation by approx. 44% |
| KV Cache Scaling (Large Context) | Ultra-low footprint (MLA compressed) | Standard footprint (GQA) | Standard footprint | Xing4.0 consumes far less KV cache VRAM at deep 128K to 256K context |
Tiel seem still very op
And Cyber-Tiel is practically the same but Abliterated; I also have it under the MiniPlus and NanoPlus recipes if you want to try them.
Hello, thanks for your work! Have you tested your mini models at 512kโ1M context, especially in llama.cpp? Does the quantization remain stable or accurate that deep into the context, or is 160kโ256k currently the validated range?
Hello! Honestly, I have only tested and validated them up to 256K tokens, which is the range where I can verify stability and accuracy. Beyond that (512K to 1M), I haven't tested them directly.
Thanks for your work IsValorum, fan of your VAL-APEX Mini and Nano. I would really like to see https://huggingface.co/agentionai/Qwen3.8-27B-AP-GGUF with VAL-APEX Abliterated. Reason being is that it requires fewer thinking tokens for similar output, also outperforms ISTA and Unsloth. Agention seems to be among the models for its size I've come across. It also has adaptive speculative decoding. Would really like it to be abliterated and with VAL-Apex applied to see it be reduced in size further with it being uncensored.
Here is the full report from his work with substantiation of why I'd like specifically this model VAL-APEX abliterated: https://share.fieldworkledger.com/tTkQDHijE_Anj0yv79bsIw
Hi @smolboon ,
Thanks a lot for the support on the Mini and Nano builds!
I checked out the Fieldwork Ledger report you linked. It is actually a benchmark comparing different GGUF quantizers (Agention vs Unsloth, ISTA, etc.) on stock Qwen3.8-27B, rather than a new fine-tuned model.
Because Agention only released final .gguf files and there are no unquantized BF16 weights, it is physically impossible to abliterate their files (abliteration has to be calculated directly on raw 16-bit weights before quantizing).
That said, we already have models in the catalog that cover exactly what you are looking for:
- If you want an uncensored model that uses significantly fewer thinking tokens, check out Qwen3.8-27B-EfficientThink-Uncensored (available in both MiniPlus and NanoPlus). It was tuned specifically to cut down thinking bloat while keeping full reasoning.
- If you just want the stock 27B base fully abliterated, Huihui-Qwen3.8-27B-Abliterated is already available.
- Plus, our MiniPlus and NanoPlus V3 builds already include the native MTP speculative decoding head directly in the GGUF.
Give EfficientThink a spin, it is probably the exact fit for what you have in mind!
Good catch @nmh2k3 , and thanks for the screenshot!
That happened because the main GGUF has the standard 64 layers (0-63) while the MTP head was packaged separately in the draft file, but the conversion script left block_count = 65 in the header. llama.cpp saw that 65 and started looking for blk.64.* inside the main file, which obviously wasn't there.
I just pushed an update to both NanoPlus V3 and MiniPlus V3 fixing the header to block_count = 64.
If you already downloaded the ~11 GB file and don't feel like downloading it all over again, you can patch your local file in 2 seconds with this quick Python snippet:
import struct
filename = "Qwen3.8-27B-EfficientThink-Uncensored-VAL-APEX-I-NanoPlus-V3.gguf"
with open(filename, "r+b") as f:
header = f.read(2048)
idx = header.find(b"qwen35.block_count")
if idx != -1:
f.seek(idx + len(b"qwen35.block_count") + 4)
f.write(struct.pack("<I", 64))
print("Done! Fixed block_count to 64.")
After that it will load right away. And if you want to run MTP speculative decoding with it, just add --model-draft mtp-Qwen3.8-27B-Q4_0.gguf to your command!
I also got an error when using the model in MLX-Serve which is my go-to inference engine atm on Mac, likely TensorFold soon tho when their app drops.
Hi @smolboon !
Did you download the file before the hotfix I pushed earlier today?
If so, you likely ran into the exact same issue as @nmh2k3 above. The GGUF header had block_count = 65 instead of 64, which causes GGUF loaders (including mlx-serve's GGUF engine) to panic looking for layer 64 (blk.64.*). If that's what happened, running that quick 3-line Python snippet above on your local file will fix it immediately without having to redownload.
Also, just in case, make sure you're passing the --mlx-gguf flag when launching mlx-serve with single-file GGUFs on Apple Silicon.
If it's a different error entirely, could you paste the terminal output / log here? Happy to check it out!
After, I use MLX-Serve, it was just a repetitive output. I already deleted the model atm. I'll do it tomorrow when I get back from work. Thanks for your work! P.S. curious to your VAL-APEX version of DavidAU's Turbo Brilliance smallest model and it's performance benchmarking
Hi,
Thankyou for your great work for the custom-tensor quantization. It worked great so far for CyberTiel MiniPlus I'm using for long coding, it utilizes my GPU more unlike other CyberTiel quants and i can upped my context more.
Would you consider to adding Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated ? I think it has better if not fully uncensored version than the base (which is not uncensored and only trained on their uncensored base) or previous version that still introduced some rejections.
What do you think? Again, thanks for all your great work and time, much appreciated.
Hi @xtronema ,
Regarding huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated: yes, I will be quantizing it into both VAL-APEX-I-NanoPlus and VAL-APEX-I-MiniPlus-V3 builds with modular MTP support.
As for LFM2.5-24B-A2B QAD, that checkpoint is trained specifically to mitigate quantization loss directly on a classic, uniform Q4_0 format rather than standard modern quants like Q4_K_M or Q4_K_S. Because it was released strictly as a pre-compiled Q4_0 GGUF without raw BF16 weights, it cannot be adapted into custom mixed-precision recipes without destructive double quantization; furthermore, our existing APEX-I MiniPlus and NanoPlus V2.1 builds already achieve the overall perceived output quality of a full Q6_K while preserving key attention and state projections at higher precision.
Hello,
Ah I see, also I didn't realize it was mostly basic q4_0 haha. Regarding your builds, yeah I agree, it feels smart enough for me for my coding, at least it's CyberTiel in my use case. And that's nice to know if it can achieve comparable performance to q6_K which already close enough to q8_0 but without the cost, that's good enough for me.
Anyway, Thanks, IsValorum for the quick response and your great work, cheers.
Are you interested in qwen 3.8 flash next nanoplus?
I've tried qwen 3.8 fn iq2 xs gsq gco via strata (because i don't have enough 64gb ram for q3)
But the results seem... pretty bad compare to qwen 3.8 27b nanoplus or iq3s gsq.
I don't know it was because of iq2 weakness or it's actually whole 3.8 flashnext's problem.
Hi @nmh2k3 ,
Two things to separate: what the architecture does, and what the quantizer chose. Both models are mixtures of experts, so the comparison is like for like.
On Flash-Next, 96 percent of the weights sit in routed experts and 1.95 percent of them are read on any given token. There is almost unlimited room to cut there at no decode cost, and a budget search always finds it. On the 27B that same figure reads zero: no experts, every weight read on every token, nothing to hide the cuts behind, so the search has to take the bits out of the dense path itself. That is what the MoE rule forbids, and it is the whole reason a hand recipe earns its place on the 27B and not here.
Now, what happens if I port my MoE recipe to the Flash-Next tensor map:
| class | my recipe, ported to Flash-Next | GSQ-RCO IQ3_S | GSQ-RCO IQ2_XS |
|---|---|---|---|
| router, all norms | F32, never quantized | F32 and BF16, never quantized | same |
| hyper-connections, indexer, PLE | BF16 and F32 | BF16 and F32 | same |
| ssm_alpha, ssm_beta | BF16 | BF16 | BF16 |
| shared expert gate and up | Q4_K | Q4_K, IQ4_XS, Q6_K | IQ4_XS, IQ3_S |
| shared expert down | IQ4_NL | IQ4_NL, Q8_0 | IQ4_NL, Q2_0 |
| routed experts gate and up | IQ3_XXS, all 48 | IQ2_S x20, IQ3_XXS x17, IQ3_S x10 | IQ2_S, IQ2_XXS, IQ1_M |
| routed experts down | IQ4_NL, all 48 | IQ4_NL x39, Q2_0 x9 | Q2_0 |
| attention gate | Q6_K | Q6_K, Q4_K, IQ4_XS | IQ4_XS, IQ3_S |
| attention qkv, q, k, v | Q4_K | Q6_K, Q5_K, Q4_K | IQ4_XS |
| attention output | Q5_K | Q6_K, Q4_K | IQ4_XS |
| ssm_out | Q5_K | Q6_K, Q5_K, Q4_K | IQ4_XS |
| token_embd | IQ3_S | IQ4_XS | IQ4_XS |
| shard 1, resident | 57.74 GB | 54.81 GB | 39.21 GB |
| whole file | 86.54 GB | 83.61 GB | 68.02 GB |
Every structural decision lands in the same place. Norms, router, hyper-connections, indexer, PLE, and both ssm_alpha and ssm_beta at BF16 on both sides. And both down projections land on IQ4_NL: mine because that tensor is 640 wide and rejects every rung of my ladder, theirs because the search chose it.
The one real difference is the expert gate and up. Mine is flat IQ3_XXS on all 48 blocks; theirs drops 20 of those 48 to IQ2_S. On the 27B the same difference is cheap, because there the ladder is 64 percent of the parameters. Here it costs 5.35 percent, because the ladder is 96 percent of the file and the price of a flat floor is paid almost in full. At the IQ3_S footprint there is no room to add it, and that is why I do not ship a NanoPlus for this model.
About the low tiers. To compete with IQ2_XS I would have to build a MicroPlus, and it is not worth your bandwidth. Put the dense path at IQ4_XS, the cheapest my rule permits there, and at their 39.21 GB there are 36.6 GB left for 120.8B of expert parameters. That is 2.42 bits per weight, and 2.42 on that ladder is IQ2_S mixed with IQ2_XXS and IQ1_M. That is their IQ2_XS build, tensor for tensor. A MicroPlus would be a rename.
Run IQ3_S. That is where the shape matches.
IsValorum
Hello. Just found your models and I must say they are very impressive and I wanted to ask something, for context I am trying to build a harness that is essentially capable of frontier intelligence and work but using local small models (Astra/fable level performance in 8B dense to 40B MOE). I truly believe and have always believed that the AI industry's current approach of creating an omniscient model that knows everything and can zeroshot everything is completely wrong. If we had focused on creating models that is not all-knowing but rather has common sense and capable of gathering required context during inference, with deeper level of understanding for a specific harness and environment. I understand such model will not work with all harnesses but this will allow us saturate the weights for general intelligence and in combination with the tools, harnesses and especially the internet could in theory be able to perform as well as or even better than frontier models. At the cost of more tokens generated which is a non-issue as it is fully intended to be used for local generation. Now I wanted to ask you,, do you know of such models right now or any specific finetunes that is utilizing this approach?
Hi @Vexdeathgrip ,
Your architectural premise is sound: rather than forcing a small model to act as an all-knowing encyclopedia, dedicating model capacity to reasoning, tool orchestration and dynamic context gathering yields far higher operational utility in an agentic harness.
If you are looking for specialized checkpoints within that 8B to 35B/40B tier designed for targeted domains, here are the primary models in this release series:
Coding and Agentic Execution:
- Occamy-1.0-APEX-I-NanoPlus-GGUF (or MiniPlus-V2.1): Architecture-level reasoning, long-context tracking under DeltaNet, complex debugging and multi-file repo management.
- Tiel-Coder-35B-A3B-APEX-I-NanoPlus-GGUF: Specialized syntax precision, terminal-bench performance and fast tool-calling loops.
Deep Web Search and Information Retrieval:
- T-Search-APEX-I-MiniPlus-V2.1-GGUF: Fine-tuned specifically for query synthesis, web research, multi-source retrieval and dense context extraction.
- Nex-N2.5-mini-APEX-I-MiniPlus-V2.1-GGUF: Highly efficient compact MoE tailored for search workflows.
Legal and Statutory Reasoning:
- Thomson-1.0-Small-APEX-I-MiniPlus-V2.1-GGUF: Domain-specific fine-tune for legal analysis, regulatory compliance and contract interpretation.
Hello!
I recently found your quants and I'd like to thank you so much for your work, your models now replaced other quants for me and the quality is amazing :)
I would like to request either the Ling tiny model or Mellum thinking model.
I know they don't demand much hardware but this exactly why I am requesting them, because with smaller sizes and good quality I can spawn more then one of them or run them on smaller edge like devices where they can actually do real work and do it fast (especially in a good enough harness and clear instructions) and they can be like fast workers for small tasks or even run them on an iGPU or CPU if I need my dGPU for non-inference work.
Thank you again for the already released models and have a nice day :)
Hello.
I had all but lost hope on local AI until I ran across your quants. You'll be happy to know you got an additional fan in me.
I am a law student with no formal education in AI but I am an AI enthusiast.
Unfortunately as a struggling student, I can only afford a 16 GB VRAM graphics card. So naturally I would like to request a nano version of your T-search and Thomson quants.
Thank you again for your work, I shall donate to the fund when I next get paid.
Hi @musharc ,
Thank you for the kind words and support!
The NanoPlus-V2.1 builds for both T-Search and Thomson-1.0-Small are already in the pipeline and on their way.
Stay tuned, they will be uploaded soon!
Hi @musharc ,
Update: Both Thomson-1.0-Small-APEX-I-NanoPlus and T-Search-APEX-I-NanoPlus are now completed and live!
You can find them here:
- Thomson-1.0-Small-APEX-I-NanoPlus-GGUF (approx. 12.55 GB)
- T-Search-APEX-I-NanoPlus-GGUF (approx. 13.02 GB, includes mmproj-Q8_0)
Both will fit comfortably within your 16 GB VRAM setup. Enjoy!
Wow that was very quick. Thank you so much!
Title: Request: Handcrafted APEX-I-MiniPlus V2.1 for Sage-A1.5 (35B MoE / 3B active)
Hi IsValorum,
I'd like to request a handcrafted APEX-I-MiniPlus V2.1 quantization for Sage-A1.5.
- Model card: https://huggingface.co/InternScience/Sage-A1.5
- Paper: https://huggingface.co/papers/2610.28578
Full benchmark context
Sage-A1.5 is a 35B MoE with only ~3B active parameters per token, purpose-built for long-horizon agentic orchestration and verification. Its scores are competitive with vastly larger frontier models on the benchmarks that matter most for quantization sensitivity:
| Benchmark | Sage-A1.5 (35B/3B) | Ornith-1.5 (35B-A3B) | Ornith-1.5 (397B) | Kimi-K3 (2.8T/104B) | GLM-5.3 (open weights) |
|---|---|---|---|---|---|
| BrowseComp | 81.7% | 67.6% | 86.6% | 91.2% | ~68.6 (index) |
| SWE-Bench Pro | 61.2% | 59.6% | 65.1% | 63.3% | โ (see DeepSWE) |
| GAIA | 95.1% | โ | โ | 78.3% | โ |
Sage-A1.5's 95.1% on GAIA is particularly notableโit outperforms substantially larger models on this long-horizon agentic benchmark. On SWE-Bench Pro, its 61.2% is within striking distance of Kimi-K3's 63.3% despite being ~35ร smaller in total parameters and ~35ร smaller in active parameters. On BrowseComp, its 81.7% surpasses the 35B-A3B Ornith-1.5 variant (67.6%) by a wide margin, though it trails the 397B Ornith flagship (86.6%) and the 2.8T Kimi-K3 (91.2%).
Ornith-1.5-35B-A3B is the closest architectural peerโsame 35B total / 3B active MoE footprint. Sage-A1.5 leads it on BrowseComp (81.7 vs. 67.6) and SWE-Bench Pro (61.2 vs. 59.6), which makes Sage-A1.5 the stronger 3B-active agentic model in this class. That is exactly why a handcrafted quantization matters: there is no slack to absorb router drift or attention degradation.
Why the handcrafted APEX-I treatment is the right fit
Community GGUF quantizations for Sage-A1.5 already circulate, but there is no handcrafted APEX-I option. The MiniPlus V2.1 recipe has a demonstrated track record on comparable MoE architectures: prior releases have held ฮPPL โ +0.05% to +1.08% while preserving the three failure points where generic quantizations break MoE reasoning:
- Routers
gate_inp/gate_shexp: uncompressed F32 โ zero routing drift - Output head: Q6_K
- Attention gates: Q8_0
- Core experts: IQ3_XXS / IQ3_S or above
Sage-A1.5's agentic orchestration and verification protocol depend directly on stable routing and attention gates. A quantization that clips the 61.2 SWE-Bench Pro or 81.7 BrowseComp margins defeats the purpose of running this model instead of a smaller generalist.
Suggested starting preset
APEX-I-MiniPlus V2.1, following the standard armor above. If the 35B/3B footprint allows a tighter envelope without crossing the quality floor, an APEX-I-MiniPlus V2 variant would also be valuable for users prioritizing theoretical edge fidelity.
Use case: local multi-step coding, tool orchestration, and scientific research-exploration workflowsโexactly where Sage-A1.5's agentic margins over similarly-sized peers are the whole point.
Thanks for considering this request.
I had already looked at this model, just give me some time as I am at university and I will generate my own I-Matrix with my own corpus that I have been polishing.
Hi IsValorum.
Love your work as always!
So I stumbled on something I found quite interesting.
For those of us with budget consumer PCs, running Qwen3.8-Flash-Next (125B A6B MoE) locally is out of reach.
However, I stumbled upon this project which claims it can be used to run Qwen3.8-Flash-Next on 16 GB of VRAM and 32 GB of RAM.
I haven't tested whether it actually works or the tok/s is acceptable, but just wondering if this is something you could work with and whether you could use any of your signature techniques on it.
I will test it myself in the coming days once I get on top of some university work.
Title: Request: Handcrafted APEX-I-MiniPlus V2.1 for Sage-A1.5 (35B MoE / 3B active)
Hi IsValorum,
I'd like to request a handcrafted APEX-I-MiniPlus V2.1 quantization for Sage-A1.5.
- Model card: https://huggingface.co/InternScience/Sage-A1.5
- Paper: https://huggingface.co/papers/2610.28578
Full benchmark context
Sage-A1.5 is a 35B MoE with only ~3B active parameters per token, purpose-built for long-horizon agentic orchestration and verification. Its scores are competitive with vastly larger frontier models on the benchmarks that matter most for quantization sensitivity:
Benchmark Sage-A1.5 (35B/3B) Ornith-1.5 (35B-A3B) Ornith-1.5 (397B) Kimi-K3 (2.8T/104B) GLM-5.3 (open weights) BrowseComp 81.7% 67.6% 86.6% 91.2% ~68.6 (index) SWE-Bench Pro 61.2% 59.6% 65.1% 63.3% โ (see DeepSWE) GAIA 95.1% โ โ 78.3% โ Sage-A1.5's 95.1% on GAIA is particularly notableโit outperforms substantially larger models on this long-horizon agentic benchmark. On SWE-Bench Pro, its 61.2% is within striking distance of Kimi-K3's 63.3% despite being ~35ร smaller in total parameters and ~35ร smaller in active parameters. On BrowseComp, its 81.7% surpasses the 35B-A3B Ornith-1.5 variant (67.6%) by a wide margin, though it trails the 397B Ornith flagship (86.6%) and the 2.8T Kimi-K3 (91.2%).
Ornith-1.5-35B-A3B is the closest architectural peerโsame 35B total / 3B active MoE footprint. Sage-A1.5 leads it on BrowseComp (81.7 vs. 67.6) and SWE-Bench Pro (61.2 vs. 59.6), which makes Sage-A1.5 the stronger 3B-active agentic model in this class. That is exactly why a handcrafted quantization matters: there is no slack to absorb router drift or attention degradation.
Why the handcrafted APEX-I treatment is the right fit
Community GGUF quantizations for Sage-A1.5 already circulate, but there is no handcrafted APEX-I option. The MiniPlus V2.1 recipe has a demonstrated track record on comparable MoE architectures: prior releases have held ฮPPL โ +0.05% to +1.08% while preserving the three failure points where generic quantizations break MoE reasoning:
- Routers
gate_inp/gate_shexp: uncompressed F32 โ zero routing drift- Output head: Q6_K
- Attention gates: Q8_0
- Core experts: IQ3_XXS / IQ3_S or above
Sage-A1.5's agentic orchestration and verification protocol depend directly on stable routing and attention gates. A quantization that clips the 61.2 SWE-Bench Pro or 81.7 BrowseComp margins defeats the purpose of running this model instead of a smaller generalist.
Suggested starting preset
APEX-I-MiniPlus V2.1, following the standard armor above. If the 35B/3B footprint allows a tighter envelope without crossing the quality floor, an APEX-I-MiniPlus V2 variant would also be valuable for users prioritizing theoretical edge fidelity.
Use case: local multi-step coding, tool orchestration, and scientific research-exploration workflowsโexactly where Sage-A1.5's agentic margins over similarly-sized peers are the whole point.
Thanks for considering this request.
The MiniPlus and NanoPlus V2.1 versions are now ready and uploaded along with the 1 Million token I-Matrix that it generates.
https://huggingface.co/IsValorum/Sage-A1.5-APEX-I-MiniPlus-V2.1-GGUF
https://huggingface.co/IsValorum/Sage-A1.5-APEX-I-NanoPlus-V2.1-GGUF
https://huggingface.co/IsValorum/Sage-A1.5-VAL-1M-I-Matrix-GGUF
Hi IsValorum.
Love your work as always!So I stumbled on something I found quite interesting.
For those of us with budget consumer PCs, running Qwen3.8-Flash-Next (125B A6B MoE) locally is out of reach.
However, I stumbled upon this project which claims it can be used to run Qwen3.8-Flash-Next on 16 GB of VRAM and 32 GB of RAM.
I haven't tested whether it actually works or the tok/s is acceptable, but just wondering if this is something you could work with and whether you could use any of your signature techniques on it.
I will test it myself in the coming days once I get on top of some university work.
I'm familiar with Strata Engine; it allows Qwen3.8 Flash Next to be used on graphics cards with up to 12GB of VRAM. In that specific case, the GSQ-RCO cuats (IQ3_S, IQ2_XXS etc) are really good, unlike the GSQ-RCO cards on the 27B.
You have many models so which one is the model that you use the most
Hi IsValorum.
Love your work as always!So I stumbled on something I found quite interesting.
For those of us with budget consumer PCs, running Qwen3.8-Flash-Next (125B A6B MoE) locally is out of reach.
However, I stumbled upon this project which claims it can be used to run Qwen3.8-Flash-Next on 16 GB of VRAM and 32 GB of RAM.
I haven't tested whether it actually works or the tok/s is acceptable, but just wondering if this is something you could work with and whether you could use any of your signature techniques on it.
I will test it myself in the coming days once I get on top of some university work.I'm familiar with Strata Engine; it allows Qwen3.8 Flash Next to be used on graphics cards with up to 12GB of VRAM. In that specific case, the GSQ-RCO cuats (IQ3_S, IQ2_XXS etc) are really good, unlike the GSQ-RCO cards on the 27B.
Would it be possible for you to do your VAL-APEX or APEX-I thing Qwen3.8-Flash-Next such that it is improved in some way when run on Strata Engine? If so, would there be any value in doing so?
Excuse my ignorance, I got no idea what those are. All I know is they give me awesome results on my budget hardware.
Hi IsValorum,
Thanks for the Sage-A1.5-APEX-I-MiniPlus-V2.1-GGUF release โ really appreciate you putting it together so quickly.
Iโve tested it on llama.cpp๏ผb11669๏ผ. The test was run through the DeepSeek harness, using the โpelican riding a bicycleโ prompt.
Test setup:
- GPU: RTX 4060 (8 GB), VRAM usage ~7.2 GB
- RAM: DDR4 32 GB
- OS: Windows 10
- Context: 131,072
- MTP: not enabled
- Test harness: DeepSeek harness
- Prompt: pelican riding a bicycle
Key launch args:
ngl=999
ncmoe=40
c=131072
ctk=q8_0
ctv=q8_0
kv-unified=true
fa=on
b=2048
ub=2048
moe-cache-mib=3072
spec-type=ngram-map-k4v,ngram-mod
spec-draft-n-max=2
spec-ngram-mod-n-match=24
spec-ngram-mod-n-min=48
spec-ngram-mod-n-max=64
temp=0.65
top-p=0.95
top-k=20
min-p=0.05
repeat-penalty=1.05
reasoning=on
reasoning-effort=xhigh
reasoning-budget=32768
Observed speeds from the log:
| Run | Tokens | Speed |
|---|---|---|
| Long run | 3,084 eval tokens | 34.18 t/s |
| Short run 1 | 127 tokens | 37.45 t/s |
| Short run 2 | 485 tokens | 37.19 t/s |
Prompt eval was around 69 t/s on the small prompt, and instantaneous tg stayed roughly 28โ34 t/s, with short 3s peaks up to ~41 t/s. The long run finished at 34.15 t/s over 3,047 generated tokens.
So on a much smaller 8 GB card itโs still quite usable, and VRAM stayed around 7.2 GB with 128K context and ncmoe=40. The pelican-on-bicycle prompt ran through fine in the DeepSeek harness as a quick sanity check.
Thanks again!
If you want even lower PCIe memory bus pressure and slightly faster inference throughput on an 8 GB VRAM setup, feel free to try the Sage-A1.5-APEX-I-NanoPlus-V2.1-GGUF variant. It provides a reduced memory footprint (12.14 GB vs 14.18 GB) while preserving virtually identical reasoning fidelity.
Surprised no Gemma4 26B in the MoE section... Any reasons for that? For language task it's still king
Please quantize this: https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
A new 35B A3B model that might ACTUALLY beat Qwen3.6-35B A3B CONSISTENTLY (nearly) on every benchmarks, but it's early and no quants of it are there yet :) your chance.
Please quantize this: https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
A new 35B A3B model that might ACTUALLY beat Qwen3.6-35B A3B CONSISTENTLY (nearly) on every benchmarks, but it's early and no quants of it are there yet :) your chance.
Iโm already preparing everything for this model; Iโll upload a precompiled version for llama.cpp (since itโs not compatible with the vanilla version), along with the BF16 weights, a Q8_0 quantization, an I-Matrix (based on 1.25M tokens), and my "Mini" and "Nano Plus" quantizations.
Is there any 35b model can beat tiel coder on coding tasks yet ?
Or tiel still the best among 35b ?
From my personal testing of MoE models I am tempted to say no.
Occamy, Kat and Ornith are viable options, though. Each in their subdomain.
IsValorum will have a more sophisticated answer to that.
Please quantize this: https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview
A new 35B A3B model that might ACTUALLY beat Qwen3.6-35B A3B CONSISTENTLY (nearly) on every benchmarks, but it's early and no quants of it are there yet :) your chance.
Iโm already preparing everything for this model; Iโll upload a precompiled version for
llama.cpp(since itโs not compatible with the vanilla version), along with the BF16 weights, a Q8_0 quantization, an I-Matrix (based on 1.25M tokens), and my "Mini" and "Nano Plus" quantizations.
THANKS DUDE YOU'RE FRICKIN' AMAZING, MY 3050 4GB WILL THANK YOU!!!
๐ฅTiel-Coder-35B-A3B-APEX-MiniPlus/NanoPlus
๐ฅOccamy1.0-35B-A3B-APEX-Mini/NanoPlus
๐ฅ Ornith1.5-APEX-Mini/NanoPlus
4๏ธโฃ KAT-Coder-Mini/NanoPlus
5๏ธโฃQwen3.8-35B-A3B-Distill Mini/NanoPlus
6๏ธโฃNex-mini-APEX Mini/NanoPlus
While Tiel Coder is beautiful, it's unbeatable. The closest is Occamy, which is very similar but of slightly lower quality. Next is Ornith, which is just as good as Tiel Coder, but...
It overthinks to the point of tripling the tokens generated simply by considering things already discarded.
Make Apex for a Reap ?
Like enable dlss performance on 360p ?
Pleeaasseee do AtomicChat/MiniCPM-V-4.7-35B-A3B-Uncensored-GGUF like you did MiniCPM V4.7
Make Apex for a Reap ?
Like enable dlss performance on 360p ?
Qwen3.88 Flash Next REAP?
Is it worth doing Mini/NanoPlus for https://huggingface.co/GrootTeam/Groot-19B-A3B-261023 ?
Iโm researching this model to make sure it isnโt another "miraculous" LFM2.5-24B-A2B that performs terribly in reality.
maybe there is another uncensored version with the BF16 version.
also, do distributedcognition/MiniCPM-V-4.7-35B-A3B-abliterated?
KRAFTON/Raon-1-mini-preview
and future versions and abliterations please
inception42/Jais-3-MoVA-36B-A4B
the only Arabic focused model which is in par with something like Gemma-4-31B, it would be great to see a MiniPlus or a NanoPlus version, the available options are still very heavy weights.
inception42/Jais-3-MoVA-36B-A4B
the only Arabic focused model which is in par with something like Gemma-4-31B, it would be great to see a MiniPlus or a NanoPlus version, the available options are still very heavy weights.
I am not quantizing MoVA; the K2 Horizon architecture is extremely KV-cache-intensiveโyouโll be lucky to get 1K of context with this model. If you want to interact with the model in Arabic, I can recommend some model for which I generated a custom I-Matrix, using a corpus specifically curated for the Arabic language.
https://huggingface.co/IsValorum/Clover-35B-A3B-Preview-APEX-I-MiniPlus-V2.1-GGUF
https://huggingface.co/IsValorum/Clover-35B-A3B-Preview-APEX-I-NanoPlus-V2.1-GGUF
inception42/Jais-3-MoVA-36B-A4B
the only Arabic focused model which is in par with something like Gemma-4-31B, it would be great to see a MiniPlus or a NanoPlus version, the available options are still very heavy weights.
And Sage (Arabic calibrated I-Matrix ) https://huggingface.co/IsValorum/Sage-A1.5-APEX-I-MiniPlus-V2.1-GGUF
https://huggingface.co/IsValorum/Sage-A1.5-APEX-I-NanoPlus-V2.1-GGUF
also, do distributedcognition/MiniCPM-V-4.7-35B-A3B-abliterated?
Quantizations uploads
https://huggingface.co/IsValorum/MiniCPM-V-4.7-35B-A3B-Abliterated-APEX-I-MiniPlus-V2.1-GGUF
https://huggingface.co/IsValorum/MiniCPM-V-4.7-35B-A3B-Abliterated-APEX-I-NanoPlus-GGUF
Axion1-0/Axion-1.0-35B next candidate for miniplus?
Axion1-0/Axion-1.0-35B next candidate for miniplus?
Yesterday I didn't have enough time to generate an I-Matrix and quantize Mini/Nano Plus, but today I do.
๐ฅTiel-Coder-35B-A3B-APEX-MiniPlus/NanoPlus
๐ฅOccamy1.0-35B-A3B-APEX-Mini/NanoPlus
๐ฅ Ornith1.5-APEX-Mini/NanoPlus
4๏ธโฃ KAT-Coder-Mini/NanoPlus
5๏ธโฃQwen3.8-35B-A3B-Distill Mini/NanoPlus
6๏ธโฃNex-mini-APEX Mini/NanoPlusWhile Tiel Coder is beautiful, it's unbeatable. The closest is Occamy, which is very similar but of slightly lower quality. Next is Ornith, which is just as good as Tiel Coder, but...
It overthinks to the point of tripling the tokens generated simply by considering things already discarded.
Isn't Cyber Tiel, better then Tiel according to it's makers
@IsValorum feel like working your quant magic on a Gemma4 or a Nemotron MoE?
ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B (or original)
ThinkingCap-gemma-4-26B-A4B-it (or original or og-QAT or QAT abliterated)
My reasons for asking is that Gemma is great at languages (I use E4B to proof OCR output) and chatting, and Nemotron Lightning is (supposedly) a sipper where it comes to context memory use. I doubt either would touch the Tiels though where it comes to actual work, so I won't be surprised if you say nope. ๐
Title: Request: Handcrafted APEX-I-MiniPlus V2.1 for Apex-35B-V1.0 (35B MoE / 3B active)
Hi IsValorum,
I'd like to request a handcrafted APEX-I-MiniPlus V2.1 quantization for Apex-35B-V1.0 by Apex Intelligence.
- Model card: https://huggingface.co/Apexintelligence-AI/Apex-35B-V1.0
- Official GGUF: https://huggingface.co/Apexintelligence-AI/Apex-35B-V1.0-GGUF
Full benchmark context
Apex-35B-V1.0 is a 35B MoE with ~3B active parameters, built on Qwen3.6-35B-A3B and further trained via a recursive self-improvement pipeline. The official card publishes a side-by-side table against Qwen3.6-35B-A3B (35B), Recursive Flint (40B), GLM-5.3 (753B), GLM-5.3-Flash (320B), and DeepSeek-V4.1-Flash (552B+196B).
Agentic / tool-use benchmarks:
| Benchmark | Apex-35B-V1.0 | Qwen3.6-35B-A3B | Recursive Flint | GLM-5.3 (753B) |
|---|---|---|---|---|
| BrowseComp | 69.4 | 42.1 | 23.2 | 80.9 |
| BFCL V4 | 84.0 | 75.2 | 65.4 | 75.8 |
| WideSearch | 69.9 | 60.9 | 48.3 | 78.4 |
| APEX-Agents | 64.6 | 41.7 | 33.8 | 68.2 |
| SWE-bench Pro v2 | 92.4 | 72.4 | 85.2 | 95.8 |
| SWE-Atlas QnA | 33.1 | 12.1 | 16.1 | 48.7 |
Research benchmarks:
| Benchmark | Apex-35B-V1.0 | Qwen3.6-35B-A3B | Recursive Flint | GLM-5.3 (753B) |
|---|---|---|---|---|
| FrontierScience | 40.0 | 21.7 | 5.0 | 30.7 |
| PaperBench | 60.9 | 45.8 | ~29 | 42.5 |
| SciCode (Main/Sub) | 62.6 / 75.6 | 37.5 / 60.6 | 26.6 / 46.0 | 79.2 / 91.3 |
| ASI-Bench | 22.5 | 16.3 | 17.5 | 33.1 |
Sources: official HF model card benchmark table; PaperBench and FrontierScience comparisons also reported in Chinese coverage. GAIA is not reported in the official benchmark table, so it is omitted here rather than inferred.
The pattern is clear: Apex-35B-V1.0 is a 3B-active model that trades blows with 300Bโ750B class systems on agentic and research tasksโBrowseComp 69.4 against GLM-5.3's 80.9, SWE-bench Pro v2 92.4 against GLM-5.3's 95.8, FrontierScience 40.0 against GLM-5.3's 30.7. That margin is fragile. It lives or dies on routing stability and attention-gate fidelity.
Why this model specifically needs the handcrafted APEX-I treatment
A community GGUF release for Apex-35B-V1.0 already exists (Sangreal ICE), and its author documents a critical property: this base model is unusually sensitive to rounding. Even a plain Q8_0 sits roughly 0.015 KLD away from BF16 on code, which is higher than other 35B-A3B ladders at the same size. The ICE 2.0 12G tier still shows 76.4% same-top-1 on wiki and 86.6โ87.9% on codeโmeaning even at 12 GB, roughly 1 in 5 wiki token distributions diverges from the BF16 master.
That sensitivity is not a reason to skip APEX-I. It is the reason APEX-I is the right treatment. The MiniPlus recipe protects exactly the tensors this model cannot afford to lose:
- Routers
gate_inp/gate_shexp: uncompressed F32 โ zero routing drift - Output head: Q6_K
- Attention gates: Q8_0
- Core experts: IQ3_XXS / IQ3_S or above
Prior handcrafted APEX-I releases on comparable MoE architectures have held ฮPPL โ +0.05% to +1.08% while preserving those three failure points. For a model whose BrowseComp and SWE-bench Pro margins are the entire reason to run it locally instead of a generic 3B model, that armor is not optional.
Suggested starting preset
APEX-I-MiniPlus V2.1, following the standard armor above. Given the documented rounding sensitivity of this base model, I would especially value the handcrafted router/attention-gate treatment over any size-optimized generic tier. If the 35B/3B footprint allows, an APEX-I-MiniPlus V2 variant would also be valuable for users prioritizing edge fidelity.
Use case: local agentic coding, tool orchestration, and research-replication workflowsโexactly where SWE-bench Pro v2's 92.4 and BrowseComp's 69.4 are the scores that justify the download.
Thanks for considering this request.
@IsValorum feel like working your quant magic on a Gemma4 or a Nemotron MoE?
ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B (or original)
ThinkingCap-gemma-4-26B-A4B-it (or original or og-QAT or QAT abliterated)
My reasons for asking is that Gemma is great at languages (I use E4B to proof OCR output) and chatting, and Nemotron Lightning is (supposedly) a sipper where it comes to context memory use. I doubt either would touch the Tiels though where it comes to actual work, so I won't be surprised if you say
nope. ๐
Tomorrow, ThinkingCap-gemma-4-26B-A4B-it will be quantized; Nemotron doesn't excel at anything.
Title: Request: Handcrafted APEX-I-MiniPlus V2.1 for Apex-35B-V1.0 (35B MoE / 3B active)
Hi IsValorum,
I'd like to request a handcrafted APEX-I-MiniPlus V2.1 quantization for Apex-35B-V1.0 by Apex Intelligence.
- Model card: https://huggingface.co/Apexintelligence-AI/Apex-35B-V1.0
- Official GGUF: https://huggingface.co/Apexintelligence-AI/Apex-35B-V1.0-GGUF
Full benchmark context
Apex-35B-V1.0 is a 35B MoE with ~3B active parameters, built on Qwen3.6-35B-A3B and further trained via a recursive self-improvement pipeline. The official card publishes a side-by-side table against Qwen3.6-35B-A3B (35B), Recursive Flint (40B), GLM-5.3 (753B), GLM-5.3-Flash (320B), and DeepSeek-V4.1-Flash (552B+196B).
Agentic / tool-use benchmarks:
Benchmark Apex-35B-V1.0 Qwen3.6-35B-A3B Recursive Flint GLM-5.3 (753B) BrowseComp 69.4 42.1 23.2 80.9 BFCL V4 84.0 75.2 65.4 75.8 WideSearch 69.9 60.9 48.3 78.4 APEX-Agents 64.6 41.7 33.8 68.2 SWE-bench Pro v2 92.4 72.4 85.2 95.8 SWE-Atlas QnA 33.1 12.1 16.1 48.7 Research benchmarks:
Benchmark Apex-35B-V1.0 Qwen3.6-35B-A3B Recursive Flint GLM-5.3 (753B) FrontierScience 40.0 21.7 5.0 30.7 PaperBench 60.9 45.8 ~29 42.5 SciCode (Main/Sub) 62.6 / 75.6 37.5 / 60.6 26.6 / 46.0 79.2 / 91.3 ASI-Bench 22.5 16.3 17.5 33.1 Sources: official HF model card benchmark table; PaperBench and FrontierScience comparisons also reported in Chinese coverage. GAIA is not reported in the official benchmark table, so it is omitted here rather than inferred.
The pattern is clear: Apex-35B-V1.0 is a 3B-active model that trades blows with 300Bโ750B class systems on agentic and research tasksโBrowseComp 69.4 against GLM-5.3's 80.9, SWE-bench Pro v2 92.4 against GLM-5.3's 95.8, FrontierScience 40.0 against GLM-5.3's 30.7. That margin is fragile. It lives or dies on routing stability and attention-gate fidelity.
Why this model specifically needs the handcrafted APEX-I treatment
A community GGUF release for Apex-35B-V1.0 already exists (Sangreal ICE), and its author documents a critical property: this base model is unusually sensitive to rounding. Even a plain Q8_0 sits roughly 0.015 KLD away from BF16 on code, which is higher than other 35B-A3B ladders at the same size. The ICE 2.0 12G tier still shows 76.4% same-top-1 on wiki and 86.6โ87.9% on codeโmeaning even at 12 GB, roughly 1 in 5 wiki token distributions diverges from the BF16 master.
That sensitivity is not a reason to skip APEX-I. It is the reason APEX-I is the right treatment. The MiniPlus recipe protects exactly the tensors this model cannot afford to lose:
- Routers
gate_inp/gate_shexp: uncompressed F32 โ zero routing drift- Output head: Q6_K
- Attention gates: Q8_0
- Core experts: IQ3_XXS / IQ3_S or above
Prior handcrafted APEX-I releases on comparable MoE architectures have held ฮPPL โ +0.05% to +1.08% while preserving those three failure points. For a model whose BrowseComp and SWE-bench Pro margins are the entire reason to run it locally instead of a generic 3B model, that armor is not optional.
Suggested starting preset
APEX-I-MiniPlus V2.1, following the standard armor above. Given the documented rounding sensitivity of this base model, I would especially value the handcrafted router/attention-gate treatment over any size-optimized generic tier. If the 35B/3B footprint allows, an APEX-I-MiniPlus V2 variant would also be valuable for users prioritizing edge fidelity.
Use case: local agentic coding, tool orchestration, and research-replication workflowsโexactly where SWE-bench Pro v2's 92.4 and BrowseComp's 69.4 are the scores that justify the download.
Thanks for considering this request.
Both versions ready
https://huggingface.co/IsValorum/Apex-35B-V1.0-APEX-I-MiniPlus-V2.1-GGUF
https://huggingface.co/IsValorum/Apex-35B-V1.0-APEX-I-NanoPlus-V2.1-GGUF
Would you be open to consider quanting decision-model oriented finetunes of Gemma4 or Qwen3.x 27B/35B-A3B?
@IsValorum feel like working your quant magic on a Gemma4 or a Nemotron MoE?
ThinkingCap-NVIDIA-Nemotron-3.5-Lightning-30B-A3B (or original)
ThinkingCap-gemma-4-26B-A4B-it (or original or og-QAT or QAT abliterated)
My reasons for asking is that Gemma is great at languages (I use E4B to proof OCR output) and chatting, and Nemotron Lightning is (supposedly) a sipper where it comes to context memory use. I doubt either would touch the Tiels though where it comes to actual work, so I won't be surprised if you say
nope. ๐
Both ready:
https://huggingface.co/IsValorum/ThinkingCap-Gemma-4-26B-A4B-it-APEX-I-MiniPlus-V2.1-GGUF
https://huggingface.co/IsValorum/ThinkingCap-Gemma-4-26B-A4B-it-APEX-I-NanoPlus-V2.1-GGUF
SmallAICreator/GRAFT-50B-A3B-GGUF
SmallAICreator/GRAFT-50B-A3B-GGUF
It's a generation prior to Qwen3.5, it's Qwen3 and it performs worse than Qwen3 30B A3B despite having more parameters (40B), pruning more than 40B of parameters from the original 80B left it lobotomized.
Furthermore, there are no BF16s or safetensors anywhere; only a Q4_K_M was uploaded.
DavidAU/Qwen3.8-40B-Ultimate-Heretic-Uncensored
ItachiSan/LiquidCode-2.5-24B-A2B
based on LiquidAI/LFM2.5-24B-A2B, i already use your LFM2.5-24B-A2B-APEX-I-MiniPlus-V2.1, which is great, and the fastest model on my pc, a coding variant would be awesome .
Suggest a model for a future handcrafted MiniPlus or NanoPlus release.
Please include:
- Upstream model link
- Intended use or strengths
- Why it would be valuable at a MiniPlus or NanoPlus footprint
I will review requests based on demand, architecture, local-inference practicality, and whether the model would benefit from a custom tensor-by-tensor quantization.
A new general model built from scratch, a new 35B A3B that beats Qwen3.6 35B A3B, that ISN'T A FINETUNE.
https://huggingface.co/maxwell-model/Maxwell-V1-Flash
It looks new nice and cool, just needs your special quants :)
I need a 35b a3b that even beat qwen 3.8 27b.
Qwen 3.6 35b a3b is easy to beat

