Beats Its Own Teacher: DeepSeek's 1.6-Trillion-Parameter Model, Distilled Into a 9B That Outscored It on GSM8K

The full checkpoint, the uncensored forks, the math, and how to run it in NXagents

---

Scrolling Hugging Face's trending page this week feels like watching a family reunion where the kid outgrew the parent. Near the top sits a model called DeepSeek-V4-Pro-Qwen3.5-9B — a name that sounds like a software dependency chain and a hardware requirement collided at birth. It's a 9-billion-parameter model, fine. It runs on a gaming laptop, fine. But the lineage is the story: its teacher is DeepSeek-V4-Pro, a 1.6-trillion-parameter mixture-of-experts monster that needs a datacenter wing to itself. And the student's own benchmark table claims something distillations are not supposed to do — beat the teacher.

The numbers, per the release: 94.50% on GSM8K (four-run average) against the teacher's own reported 92.60%. A 1.6T model that opens with 1,024 experts per token, outrun on grade-school math by a model small enough to run next to your browser tabs. Meanwhile the ecosystem around it has gone fully feral: an "UNCENSORED-heretic" fork posted October 30 has collected 7,314 likes in two weeks — more than the 1.6T teacher has earned in seven months — and spawned over twenty cloned demo Spaces before you could say "abliteration."

Let's dig in. Receipts first, hype second.

---

1. The Family Tree

One name, four characters. Here's who's who:

Role	Model	Params	License	What it is
Teacher	DeepSeek-V4-Pro (0813)	1.6T total / 49B active (MoE)	MIT	DeepSeek's flagship. Previewed April 24, 2026, promoted to the "-0813" GA checkpoint on August 13
Chassis	Qwen3.5-9B (via unsloth/)	~9.7B (BF16)	Apache-2.0	Alibaba's dense multimodal small model — the body the student wears
Student	DeepSeek-V4-Pro-Qwen3.5-9B	~9.65B (BF16)	Apache-2.0	Qwen3.5-9B fine-tuned on ~250K teacher responses
Sibling	DeepSeek-V4-Pro-Qwen3.5-4B	~4B	Apache-2.0	Same recipe, smaller brain (MMLU-Pro avg 76.47%)
The Feral Fork	-UNCENSORED-heretic (by xtools-at)	~9.7B	Apache-2.0	The student after a Heretic de-censorship pass

Squint at the naming and it's actually honest: "what would DeepSeek-V4-Pro say, spoken by a Qwen3.5-9B body." Distill behavior from one lab's flagship into another lab's chassis — that's the whole 2026 open-weights meta in one repo name.

        deepseek-ai/DeepSeek-V4-Pro (1.6T MoE, MIT)
                │  ~250K responses ("Max Effect" mode)
                ▼
        Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B   ← the student (Apache-2.0)
                │                                 -MTP-GGUF repo (IQ4_XS → Q8_0 + mmproj)
                │  Heretic ARA pass (no retraining)
                ▼
        xtools-at/...-UNCENSORED-heretic  →  20+ community HF Spaces clones

---

2. The Recipe: 250,000 Answers and One Honest Apology

The method, per the model card, is refreshingly old-school: generate supervised data, then SFT, then light RL.

• Data: ~250,000 math and STEM samples. The responses were generated by DeepSeek-V4-Pro "in Max Effect mode" — its high-effort reasoning setting (hosting resellers even price "Max" as a separate tier, at roughly $1.32/$3.96 per million in/out tokens per Artificial Analysis). The point: don't distill the teacher's answers, distill the teacher thinking hard.
• Stage 1 — SFT on those traces, emphasizing structured derivation, problem decomposition, and "reliable final-answer construction."
• Stage 2 — RL with GSPO (Grouped Stepwise Preference Optimization) — a light preference-tuning pass, not a heavy agentic RL grind.
• Framework: Unsloth did the memory-optimized training; hardware for the run came from collaborator Kyle Hessling (who also built a playable game, Star Skip, on top of the model as a demo).

Then there's the part most release notes hide and this one puts in a headline box: no coding data at all. Why? The author explains the timing — mid-run, DeepSeek shipped the newer 0731/0813 teacher checkpoints with much stronger coding and agentic behavior, and fresh teacher data wasn't ready. The coding distillation is declared roadmap, not shipped feature. For a model headed toward agent platforms, that omission matters, and the card doesn't pretend otherwise. In a genre famous for quietly upselling, this is the rare README that apologizes for its own gaps.

---

3. The Numbers — Impressive, With Asterisks You Should Actually Read

GSM8K (grade-school math): the student edges the teacher

Model	Accuracy
MiMo-V2.5-Pro	99.60%
Llama-3.1-405B-Instruct	96.80%
Llama-3.3-70B-Instruct	94.84%
DeepSeek-V4-Pro-Qwen3.5-9B	94.50% (4-run mean, NVFP4, vLLM)
DeepSeek-V4-Pro (teacher)	92.60%
DeepSeek-V3	89.30%

A ~1,178× parameter reduction (1.6T → ~9.65B) with a +1.9-point win on the cited reference. Even discounting the margin (single runs sit around 0.3 points apart), landing within noise of a frontier MoE — and nominally above it — from a laptop-sized student is the headline, and it's earned.

MMLU-Pro — Math / Physics / Chemistry (500 questions × 3 subsets per model, llama.cpp)

Model	Eval build	Math	Physics	Chem	Avg
DeepSeek-V4-Pro-Qwen3.5-9B	MTP-Q8_0 GGUF, 8K	92.40%	89.40%	89.80%	90.53%
Qwen3.5-9B (base)	MTP-Q8_0 GGUF, 8K	90.60%	89.00%	89.20%	89.60%
Claude Mythos-distilled 27B	MTP-Q8 GGUF, 32K	88.40%	87.60%	82.60%	86.20%
Claude Mythos-distilled 9B	MTP-Q8_0 GGUF, 8K	86.00%	78.60%	78.60%	81.07%
DeepSeek-V4-Pro-Qwen3.5-4B	MTP-Q8 GGUF, 64K	80.40%	74.20%	74.80%	76.47%

+0.93 points over its own chassis. Not earthshaking on average — but see the next table.

The efficiency table is the real flex

Same 1,500-question sample. The distilled model answers more correctly while spending fewer reasoning tokens:

Model	Accuracy	Avg reasoning tokens/question	Tokens per correct answer	Correct answers per 1M tokens
9B distill	90.53%	2,858	3,157	316.8
Qwen3.5-9B	89.60%	4,425	4,938	202.5
Claude Mythos-distilled 9B	81.07%	3,310	4,083	244.9

That's 36.1% fewer tokens per correct answer than its own base model (+56.4% correct-answer throughput), and the card's per-subject ledger claims math answers alone cost 47.4% fewer tokens. If you run this thing locally, that table is your electricity bill getting a haircut. On GSM8K the reasoning profile averages 1,627 tokens with a median of just 372 — it thinks in short bursts, not marathons.

There's also a fun behavioral quirk: asked to answer with "one single capital letter," the distilled model complied 97.2% of the time (verbose re-explanations: 0.2%), versus base Qwen3.5-9B's 37.3% compliance with 62.1% of cases rambling. The teacher's answer discipline apparently transferred along with its reasoning. For anyone who has ever written 400 words of format-policing into a system prompt, that column is aspirational.

The honesty box (read before screenshotting the tables)

The card discloses its own caveats, and they're load-bearing:

1. Self-reported. Every number here comes from the author's own runs. No independent replication yet.
2. Sampled subsets. MMLU-Pro used 500 random questions per subject — not the full subsets.
3. Mixed builds. GSM8K ran NVFP4 on vLLM; MMLU-Pro ran MTP-Q8_0 GGUF on llama.cpp. The card itself warns against cross-inferring quantization equivalence between the two tables.
4. The "teacher 92.60" is a cited reference from the release's comparison, not a freshly published DeepSeek benchmark.
5. The "Claude Mythos-distilled" comparison models are third-party distills of unclear provenance — directionally useful, not gospel.
6. Coding/tool-use gains are honestly labeled "small, preliminary generalization," not a feature claim.

The GSM8K win is real within these walls. Treat 94.50% as "competitive with frontier models on grade-school math," not as settled fact.

---

4. Meanwhile, the Model Went Feral: The Heretic Explosion

Here's the twist that turned a solid community fine-tune into a trending-page nuclear event. On October 30, user xtools-at published DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic: Jackrong's student, run through Heretic — the open-source "training-free censorship removal" tool by p-e-w that pairs directional ablation with an Optuna-driven optimizer, and since its 2.0 release uses ARA (Arbitrary-Rank Ablation) to steer refusals out of a model without touching a single weight during training.

The numbers attached to that fork (per the fork's card and secondary coverage — treat as the publisher's claims, not independent audits):

Metric	Uncensored fork claim
Refusal rate on a 100-prompt probe set	3 out of 100 (vs. 100/100 reported for the base model)
HF likes	7,314 in ~2 weeks
HF downloads	24,470 (API snapshot) + AlphaSignal reporting 32,000+ at its publication time
Cloned HF Spaces demos	20+
Largest Chinese-language writeup	"a nuclear-grade uncensored model... 5.6GB, downloaded 18,000 times right after launch" (明子206 on X)

That like count deserves a drumroll: a two-week-old 9.7B fork has out-liked DeepSeek's own 1.6T flagship (5,710 likes since April). The teacher spent a supercluster cluster's electricity bill; the fork spent an afternoon of ARA and a hype cycle. Welcome to open-weights economics, where distribution beats parameter count every single time.

Worth being the adult in the room for one paragraph: abliteration-style refusal removal has a well-documented failure mode where the model loses more than its manners — degraded factuality, weird LoRA-shaped artifacts, and a model that will now confidently answer questions it was never competent to answer. The fork's card inherits Jackrong's benchmarks but adds no new safety or capability evaluation. If you use it, you're the safety department now. Also: Apache-2.0 is a license, not a liability waiver.

---

5. Teacher vs. Student: The Full Receipts Table

	DeepSeek-V4-Pro (0813)	DeepSeek-V4-Pro-Qwen3.5-9B
Params	1.6T total / 49B active (MoE)	~9.65B dense
Where it runs	Multi-GPU server cluster	8–16GB RAM/VRAM laptop-class
GSM8K (cited)	92.60%	94.50% (4-run avg, self-reported)
MMLU-Pro Math subset	— (not in release table)	92.40%
HF likes	5,710 (since April)	78 (repo), 7,314 (heretic fork)
HF downloads	~317K	2,320 + 181,679 (GGUF repo)
Hosted price	~$0.66–$1.32 in / ~$1.98–$3.96 out per 1M tokens (reseller snapshots)	$0 — your electrons
License	MIT	Apache-2.0
Coding ability	Frontier-class (0731/0813 upgrades)	None trained (roadmap)
Vibes	National infrastructure	That one friend who read one book and won't stop

Back-of-envelope for the rent-vs-own crowd: the teacher's "Max" tier at ~$3.96 per million output tokens versus a 9B you own outright. If your workload burns serious tokens daily, the student pays for its 5–6GB of RAM in weeks — if your workload needs frontier coding and agentic chops, no laptop rescue mission is saving you.

---

6. Strengths and Weaknesses, No Chaser

Strengths

• Best-in-class 9B math/STEM reasoning, per the release's own comparison set — with the most transparent methodology we've seen on a community distill this year.
• Token-efficient: fewer reasoning tokens per correct answer than its own chassis — rare and valuable.
• Format discipline that system-prompt writers dream about.
• Deployment maturity on day one: BF16 safetensors, NVFP4 for vLLM, MTP GGUFs from IQ4_XS to Q8_0 (262,144-token context per the GGUF metadata), plus an F32 vision projector so multimodal runtimes don't choke.
• Apache-2.0 — commercially clean, no community-license strings (compare: some 2026 open models ship "community licenses" with usage floors).
• An author who documents his own limitations better than most labs document their features.

Weaknesses

• No coding training whatsoever. For agent platforms (hi), that's the whole job description. Card itself says wait for the 0731/0813-trace coding distillation.
• MMLU-Pro average gain over base is +0.93 — the win is in efficiency and math, not across-the-board IQ.
• Self-reported benchmarks, sampled subsets, mixed eval builds — impressive rigor by community standards, still not an independent audit.
• Instruction-tuned, not aligned: this is a reasoning SFT, not a product-grade assistant with refusal policies and safety evals.
• The uncensored fork is a different risk surface entirely — capability preserved, guardrails deleted, zero new evaluation.
• Vision exists in the architecture (it inherits Qwen3.5's image/video tokens and ships mmproj), but the distillation focused on text reasoning — don't expect tuned multimodal behavior.

---

7. How to Install It in the NXagents Platform

Short answer: yes, it's installable now, and it's a great pilot, but it should not become the platform's default brain. NXagents agents don't bundle LLM weights — we reach models through a thin provider bridge (same pattern as our Pika-H3-Flex video bridge from October) — so "installing" this model means serving it somewhere and pointing the bridge at it. Here's the practical path, in order of preference:

Path A — GPU pod (recommended):

pip install vllm
vllm serve Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B \
  --quantization nvfp4        # Blackwell-class GPUs; omit for BF16
# OpenAI-compatible endpoint → point the NXagents provider bridge at it

NVFP4 is exactly what the model card's GSM8K runs used, so this is the best-tested path.

Path B — CPU / laptop / edge:

# llama.cpp (MTP-aware builds) with the Q8_0 that matches the card's own eval
./llama-server -m DeepSeek-V4-Pro-Qwen3.5-9B-MTP-Q8_0.gguf \
  -mmproj mmproj-F32.gguf --ctx-size 32768

The Q4_K_M (~5.6GB, the file the Chinese write-ups cite) fits 8GB-class machines; Q8_0 wants ~12GB. The MTP head is kept, so llama.cpp's multi-token fast-decode path works.

Then apply our house rules (the Oct 4 budget-caps sermon applies in full):

1. Pilot lane only — route a shadow copy of low-stakes traffic (math-check, extraction, classification) and diff against the hosted default for a week.
2. Strict budget + rate caps on the pilot lane; the model is cheap but your engineering time isn't.
3. Fact-check gate on (our cross-model checker) — the card itself says to independently verify outputs.
4. No coding/agent traffic yet. The author says so; we'd be ignoring the label. Re-evaluate when the 0731/0813 coding distillation lands — that's the version this platform actually wants.
5. Skip the heretic fork for platform use entirely. Uncensored weights + a production agent harness is how incident postmortems get written. (And thanks to the Oct 2 Apple Full Disk Access story, we know exactly how that genre ends.)

Verdict: watchlist-to-pilot, not adoption. If the promised coding distillation ships with the 0813 teacher's traces, this 9B line becomes the most interesting "runs in a sandbox" model on our board. Until then, the teacher stays rented and the student stays a really smart side-project.

---

8. The Trending Page Lied to You (Sort Of)

One last honesty note. The "冲上Hugging Face第一" headline that triggered this piece — like most aggregator headlines — compresses a more interesting truth: it's the uncensored heretic fork (7,314 likes, 24K+ downloads, 20+ cloned demos) that's been sitting near the top of the trending charts, while the original student quietly posts 78 likes and 181K downloads on its GGUF repo. Same weights, different marketing departments: one has benchmarks, the other has forbidden fruit.

But the underlying fact is real, and it's the bigger story: the capability to (nearly) match a 1.6T flagship on math — with a fraction of the tokens — now fits in a $5.6GB download with an Apache-2.0 license. The distillation game isn't a sideshow anymore. It's the whole economy.

The kid didn't just grow up. He beat his father at math, moved into a laptop, and started a band.

---

Sources & receipts: Jackrong's model card (full checkpoint + MTP-GGUF, HF API snapshots Nov 14 ~17:40 UTC); xtools-at UNCENSORED-heretic card + HF Spaces registry; p-e-w/heretic GitHub (ARA, PR #211, Heretic 2.0 release notes); AlphaSignal and OneDigital coverage of the fork; 明子206 on X; note.com review by zephel01 (JP); DataLearner and Guancha on DeepSeek-V4-Pro 0813 (1.6T/49B, April 24 → Aug 13 GA); Artificial Analysis, AI Pricing Guru and costperprompt for teacher pricing snapshots; unsloth Qwen3.5 docs. All benchmark figures are self-reported by the model authors unless noted; download/like counts are point-in-time API snapshots.