GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
staccs
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
staccs
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
This repo is the Unicron-assimilated Qwen3.5-9B prove (4 shards + unicron reports). Apache 2.0 from Qwen3.5-9B. It is not DeepSeek Flash and not Qwen 397B.
| Job | State |
|---|---|
| This 9B assimilate upload | on the card (weights in-repo) |
| Qwen3.5-9B full assimilate (RTX 4090, budget 0.01) | DONE — 20.2% of fp16, ppl 29.8454 → 29.8400. See "Measured runs" below. |
| Qwen3.5-72B assimilate (RTX 3060, budget 0.05) | DONE — 145.4GB → 57.8GB (39.7% of fp16), wikitext-2 ppl 15.2443 → 15.2389 (−0.04%) |
Moose 0.2.11 Qwen3.5-0.8B --registers 128 --passages 200 | DONE — published at staccs/Qwen3.5-0.8B-lecore-assimilated |
| DeepSeek-V4-Flash in-weight HRR | staccs/lecore-deepseek-v4-flash-hrr |
| OpenRouter Provider | not applied |
Both from operator hardware, not a datacentre. Numbers are read straight from
the run artifacts (qwen9b_result.json, qwen72b_report.json,
wikitext_ppl.log), not restated from memory.
| Qwen3.5-9B (RTX 4090) | Qwen3.5-72B (RTX 3060) | |
|---|---|---|
| params | 9.65B | 72B (37 shards) |
| budget | 0.01 | 0.05 |
| mean bits/weight | 3.234 | 5.038 |
| size vs fp16 | 20.21% | 39.75% (145.4GB → 57.8GB) |
| perplexity before → after | 29.8454 → 29.8400 | 15.2443 → 15.2389 |
| delta |
Both deltas are negative — perplexity came out marginally lower after assimilation. That is within noise and should be read as "no measurable regression", not as evidence that assimilation improves a model.
The 72B run is the more useful datapoint: a 145GB model reduced to 57.8GB on a 3060, with wikitext-2 perplexity measured over the full test split (sliding window, 297,193 tokens) rather than a sample.
Prove run, not a capability-beat of vanilla 9B.
A Qwen3.5-9B checkpoint that went through leCore’s Unicron gate — built to sit under holographic (HRR) memory & programs.
Apache 2.0 from the base. Prove run, not a “we beat GPT” claim.
Think of a normal model as a brain with a whiteboard (context window). When the board fills, old stuff gets erased.
leCore adds a second trick: holographic memory — like writing with light through the whole crystal. Scratch half the crystal and you still read the picture. That’s HRR.
| What you care about | Normal chat model vibe | This stack (Qwen3.5-9B + leCore HRR) |
|---|---|---|
| “How much can it keep in one go?” | Many models choke at 4k–32k tokens. You paste a PDF, it forgets the start. | Base window here is 262,144 tokens native (~a whole book+). Upstream also ships YaRN ~1M extension. Not infinite — still a hard ceiling — but it’s “bring the whole thread / repo chunk / filing cabinet,” not “hope 8k fits.” |
| “What if storage / state gets wrecked?” | Normal memory is like a spreadsheet: delete 10% of cells and whole rows die. | HRR memory is damage-tolerant. Same lab test (16 facts in 1024 numbers): trash 40% of slots → holographic still 100% recall; normal packing is already ~0%. Your “notes” degrade gracefully instead of hard-crashing. |
| “Can it mix facts like Lego?” | Usually needs more training or brittle prompts. | Bind / unbind = compose roles & facts with invertible math (FHRR). Name ↔ attribute ↔ value without a mystery neural adapter. |
| Tokens (approx) | Human vibe | |
|---|---|---|
| Many “chat” models | 4k–32k | A long Discord scroll, then amnesia |
| This base (native) | 262k | Novel / big codebase slice / fat research dump in one shot |
| This base (YaRN / ext, upstream) | ~1M | Multi-book / serious long-doc — still not ∞ |
| True infinite | ❌ | Transformer KV is not infinite; HRR external memory is big + sturdy, but has a measured capacity cliff |
| Qwen3.5-9B original | This repo (Unicron-assimilated) | Factored sidecar | 27B / 397B / V4-Flash refs | |
|---|---|---|---|---|
| License | Apache 2.0 | Apache 2.0 | — | Apache / Apache / MIT |
| Context | 262k native; ~1M YaRN (upstream) | same | — | see each card |
| On-disk | 19G BF16 | 19G BF16 | skipped (no pay) |
python
from transformers import AutoModelForCausalLM, AutoTokenizerm = AutoModelForCausalLM.from_pretrained("staccs/lecore-qwen35-9b-assimilated",torch_dtype="auto",device_map="auto",)t = AutoTokenizer.from_pretrained("staccs/lecore-qwen35-9b-assimilated")
Same shapes as Qwen/Qwen3.5-9B. For HRR faculties (bind/unbind, HoloMachine, fingerprints), use leCore 0.2.11.
0.2.11 / Unicron assimilate + sample evaltests/test_degradation_table.pyNot frontier AGI. Not infinite context. Not “we compressed 9B for free.”
We measured retention, kept the weights honest, and wired them toward holographic memory you can actually explain to a human.
Capability delta vs vanilla Qwen3.5-9B: 0.00%. That is disclosed, not hidden, and it is the correct expectation for a prove run: Unicron's gate certifies and installs, it does not retrain. Anyone benchmarking this against stock Qwen should expect a tie, and independent reproduction (ratimics v6) found the same: assimilation regression 0%, no quality or speed advantage from the HRR arm.
Where the value is claimed instead — memory, not accuracy. The requantize pass took this model from 18 GiB to ~4 GiB, with every sampled tensor landing at 3 bits. On a 16 GB laptop that is the difference between "resident" and "impossible", which is the whole point for zero-egress / local-first deployments.
What is NOT yet measured, stated plainly. There is no head-to-head against the obvious
baselines — llama.cpp Q3_K_M and AWQ at matched quality — on the two numbers that decide
it: resident RSS (init and steady state) and perplexity. Until that exists, "3-bit, 4.5x
smaller" is a real measurement of this pipeline, not a demonstrated advantage over standard
quantization. Treat the memory claim as promising and unproven rather than settled.
| −0.018% |
| −0.04% |
| wall clock | 5h 10m | 2h 21m |
| tensors quantized | 369 | 560 |
| left fp16 / skipped | 0 / 3 | — / 403 |
| “Can I run little programs on the memory?” | Tool calling / agents, heavy stacks. | HoloMachine: short holographic instruction streams on the same algebra — inspectable, deterministic, no black-box weights inside the HRR core. |
| “Did you shrink the model / make it smarter?” | Marketing often lies here. | This prove: Unicron looked at every fat matrix, found heavy-tail / well-trained regime → cut nothing. Sample perplexity 2.043 → 2.043 (Δ +0.00%). Honest win = safe gate + HRR surface, not a free size drop on 9B. |
| “What’s in it for me shipping / memeing?” | Weights alone. | Drop-in Qwen3.5-9B plus a path into leCore: long-context chat, holographic scratchpad that survives corruption, composable memory ops — MIT/Apache-friendly open stack. Next sell SKU we’re aiming at is DeepSeek V4-Flash (bigger iron). |
| larger → multi-hundred-G |
| Serve VRAM (rough) | ~18–24GB BF16 + KV | same | — | Flash class ~170–175GB native |
| Unicron | — | 0 filtered, heavy-tail passthrough | — | — |
| Sample ppl | 2.043 | 2.043 (+0.00%) | — | — |