gbuzhf/Nex-N2.5-mini-OrcaRouter-Sangreal-ICE

Nex-N2.5-mini OrcaRouter — Sangreal ICE

Five ICE 2.0 tiers of nex-agi's Nex-N2.5-mini (35B total, 3B active, Qwen3.5-MoE architecture), abliterated trunk by OrcaRouter, quantized against Sangreal — a purpose-built 77-bucket calibration corpus — with an importance matrix measured on Nex itself, and shipping Qwen-Sharp v22.5.0-arrow4.0 — our agent-tuned chat template, a heavily modified derivative of peculiar-ragdoll's Qwen-Sharp (spec below). Vision works through this trunk's own projector, mmproj-Nex-N2.5-mini-OrcaRouter-Q8_0.gguf.

⚠️ Uncensored. This is an abliterated checkpoint — refusal behaviour has been suppressed in the weights. Sandbox it at the OS level and control its network and code-execution access; with no refusal backstop, a prompt injection from a hostile page or third-party code has nothing to stop it.

ICE 2.0 — arrived

All five tiers now ship as ICEv2: 12G-ICEv2, 15G-ICEv2, 19G-ICEv2, 23G-ICEv2 and 25G-ICEv2.

ICE 2.0 keeps ICE 1.6's recipe — every tensor type, the always-on path and the file size are unchanged — and adds a Sangreal-calibrated post-quantization repair of the routed experts, fitted against this trunk's own BF16. Same GGUF types, stock llama.cpp.

Paired against the same tier's ICE 1.6 — same chunks, same BF16 reference, same session (chunks better / chunks):

tier	code @2K	code @16K	wiki @2K
25G-ICEv2	-4.3% (41/64)	+0.2% (6/8)	+3.3% (34/64)
23G-ICEv2	-4.1% (43/64)	-2.8% (6/8)	+0.0% (39/64)
19G-ICEv2	-14.6% (60/64)	-6.2% (8/8)	-11.9% (58/64)
15G-ICEv2	-12.6% (63/64)	-3.7% (8/8)	-7.6% (52/64)
12G-ICEv2 †	-35.6% (64/64)	-36.0% (8/8)	-30.3% (64/64)

† 12G is new with ICE 2.0 (no ICE 1.6 12G), so it is paired against orcarouter's Q2_K, which is 1.04 GB larger.

The trunk

orcarouter/Nex-N2.5-mini-Uncensored is the abliterated (refusal-removed) build of nex-agi/Nex-N2.5-mini: a ~34B Mixture-of-Experts agentic VLM, hybrid gated-delta-net + full attention, 256 routed experts top-8 beside a shared expert on every layer, a native vision tower, and a 262K-token context. The refusal direction is baked straight into the official BF16 shards — same dtype, same shard layout, same model.safetensors.index.json, all 1026 tensors matching the base checkpoint in name, dtype and shape.

Per their card, 122 of those 1026 tensors are edited — the matrices that write the residual stream — while the MoE router, the input projections, every norm and lm_head are untouched. We verified it before reusing the calibration: by ranged-GET probe, mlp.gate.weight (the router) at layers 0, 19 and 39 is byte-identical to nex-agi's, q_proj and the shared expert's up_proj are byte-identical, and the edited writers (o_proj, linear_attn.out_proj, the shared expert's down_proj) move ~2.0–2.2% relative. Since an imatrix on a sparse MoE is mostly a statement about which experts fire, and the router is bit-for-bit the same, Sangreal's matrix — measured on the original Nex — is valid for this trunk.

At a glance — against orcarouter at the same size

Mean KLD to BF16, one session. Size-matched = orcarouter's own quant ladder for this same abliterated trunk, interpolated (log-linear) at each ICE file's exact byte size; negative = ICE closer to BF16. Paired = the same 64 chunks scored for both files against the same reference. ICEv2 rows: the 2026-10-10 session (8× RTX 3090, same binary, same reference), where every orcarouter quant and ICE 1.6 tier was re-measured; struck values are the 2026-10-04 ICE 1.6 numbers. † 12G-ICEv2 is smaller than orcarouter's smallest quant, so instead of a size-matched point it is compared directly against Q2_K (12.94 GB, 1.04 GB larger).

tier	size	code @2K	code @16K	wiki @2K	paired, code @2K (t, chunks ICE better)
12G-ICEv2	11.90 GB	-35.6% †	-36.0% †	-30.3% †	vs Q2_K (12.94 GB): -35.6% (t -17.0, 64/64)
15G-ICE 15G-ICEv2	14.51 GB	-47.4% -54.0%	-42.7% -45.4%	-45.9% -49.2%	vs Q2_K (12.94 GB): -59.8% (t -12.2, 64/64); vs Q3_K_M (16.76 GB): -22.6% (t -8.7, 64/64) vs Q2_K (12.94 GB): -64.9% (t -12.0, 64/64); vs Q3_K_M (16.76 GB): -32.1% (t -10.7, 64/64)
19G-ICE 19G-ICEv2	18.47 GB	-26.6% -39.1%	-31.5% -32.7%	-22.9% -34.5%	vs Q3_K_M (16.76 GB): -64.8% (t -9.0, 64/64); vs IQ4_XS (18.73 GB): -17.9% (t -10.5, 62/64) vs Q3_K_M (16.76 GB): -70.0% (t -9.3, 64/64); vs IQ4_XS (18.73 GB): -32.3% (t -8.1, 64/64)
23G-ICE 23G-ICEv2	22.52 GB	-31.4% -33.4%	-26.4% -30.3%	-29.9% -32.6%	vs IQ4_XS (18.73 GB): -53.6% (t -9.4, 64/64); vs Q6_K (28.51 GB): +27.3% (t +6.1, 3/64) vs IQ4_XS (18.73 GB): -55.1% (t -8.7, 64/64); vs Q6_K (28.51 GB): +24.2% (t +5.4, 8/64)
25G-ICE 25G-ICEv2	24.53 GB	-27.5% -31.5%	-20.5% -25.4%	-27.8% -29.4%	vs IQ4_XS (18.73 GB): -60.2% (t -9.6, 64/64); vs Q6_K (28.51 GB): +9.3% (t +3.6, 21/64) vs IQ4_XS (18.73 GB): -62.5% (t -8.7, 64/64); vs Q6_K (28.51 GB): +3.6% (t +1.7, 34/64)

Code at 16K context (code.test.raw, n_ctx 16384 × 8), mean KLD to BF16 vs file size; lower-left is better. All points from the 2026-10-10 session (one base, one binary); raw numbers in measurements/2026-10-10/, the 2026-10-04 ones in Table 2.

Demo — Coral Dragon Reef Rescue

A complete browser game — one self-contained index.html, no assets, no network calls — generated in a single shot by the 23G ICE 1.6 build (Nex-N2.5-mini-OrcaRouter-Sangreal-23G-ICE.gguf) in about 45 minutes, running locally in the Nous Hermes desktop harness on Windows 11 with an RTX 3070 (8 GB VRAM), 32 GB RAM and no MTP.

Play: Hugging Face Space · GitHub Pages · Source

Measurements

One binary (llama.cpp 3653e6d), one reference (this trunk's own BF16, in this repo), one session (2026-10-04, 2× RTX PRO 5000 Blackwell 48 GB). 64 chunks at n_ctx 2048; the long-context table uses 8 chunks at n_ctx 16384. Every file below was measured in this session. Raw logs: measurements/2026-10-04/. ICEv2 rows were measured on 2026-10-10 in a second session (8× RTX 3090, same binary, same BF16 reference) in which every file above was re-measured within ±4% of these values; raw logs: measurements/2026-10-10/.

Table 1 — Code (code.test.raw)

PPL(base) 2.328373 · BF16 = 100

file	size	mean KLD	99% KLD	99.9% KLD	PPL ratio	same top-1	active bpw	file bpw	overall
≈ 28.5 GB
orcarouter Q6_K	28.51 GB	0.013501	0.1844	0.9720	1.0008	97.13 %	6.754	6.581	98.2
≈ 24.5 GB
Sangreal 25G-ICEv2	24.53 GB	0.014332	0.2085	1.0195	0.9992	97.14 %	7.385	5.662	98.2
Sangreal 25G-ICE	24.53 GB	0.014750	0.2087	1.0719	1.0019	96.99 %	7.385	5.662	98.1
≈ 22.5 GB
Sangreal 23G-ICEv2	22.52 GB	0.017185	0.2662	1.2184	1.0002	96.86 %	7.215	5.197	97.9
Sangreal 23G-ICE	22.52 GB	0.017195	0.2525	1.1518	1.0022	96.64 %	7.215	5.197	97.8
≈ 18–19 GB
Sangreal 19G-ICEv2	18.47 GB	0.025892	0.4097	1.6937	1.0066	96.09 %	6.872	4.263	97.1
Sangreal 19G-ICE	18.47 GB	0.030379	0.4545	1.9283	1.0108	95.69 %	6.872	4.263	96.6
orcarouter IQ4_XS	18.73 GB	0.037021	0.5240	2.2304	1.0113	95.06 %	5.076	4.323	96.0
≈ 16.8 GB
orcarouter Q3_K_M	16.76 GB	0.086420	1.3055	4.4527	1.0426	92.38 %	4.562	3.869	92.1
≈ 14.5 GB
Sangreal 15G-ICEv2	14.51 GB	0.058596	0.9824	3.1498	1.0224	93.79 %	6.531	3.350	94.3
Sangreal 15G-ICE	14.51 GB	0.066931	1.0910	3.5109	1.0296	93.43 %	6.531	3.350	93.6
≈ 12–13 GB
Sangreal 12G-ICEv2	11.90 GB	0.107590	1.7186	5.3098	1.0525	91.86 %	6.313	2.746	90.8
orcarouter Q2_K	12.94 GB	0.166646	2.2780	5.6615	1.1055	89.07 %	3.980	2.987	86.7

Table 2 — Code at long context (code.test.raw, n_ctx 16384 × 8)

PPL(base) 2.398230 · BF16 = 100

Compare rows within this table only.

file	size	mean KLD	99% KLD	99.9% KLD	PPL ratio	same top-1	active bpw	file bpw	overall
≈ 28.5 GB
orcarouter Q6_K	28.51 GB	0.072845	1.3317	11.5045	0.9939	96.25 %	6.754	6.581	94.1
≈ 24.5 GB
Sangreal 25G-ICEv2	24.53 GB	0.068734	1.1232	11.5023	0.9997	96.36 %	7.385	5.662	94.4
Sangreal 25G-ICE	24.53 GB	0.070928	1.2198	11.9453	1.0005	96.21 %	7.385	5.662	94.2
≈ 22.5 GB
Sangreal 23G-ICEv2	22.52 GB	0.070205	1.1835	11.4419	0.9981	96.07 %	7.215	5.197	94.2
Sangreal 23G-ICE	22.52 GB	0.072753	1.2301	11.7193	0.9966	96.05 %	7.215	5.197	94.1
≈ 18–19 GB
Sangreal 19G-ICEv2	18.47 GB	0.084865	1.5289	12.5688	0.9953	95.39 %	6.872	4.263	93.1
Sangreal 19G-ICE	18.47 GB	0.086972	1.7724	11.7056	1.0068	95.08 %	6.872	4.263	92.9
orcarouter IQ4_XS	18.73 GB	0.119762	3.0800	13.2302	1.0122	94.20 %	5.076	4.323	90.8
≈ 16.8 GB
orcarouter Q3_K_M	16.76 GB	0.186525	4.9562	15.1364	1.0003	91.89 %	4.562	3.869	86.6
≈ 14.5 GB
Sangreal 15G-ICEv2	14.51 GB	0.127506	2.8844	12.8176	1.0356	93.52 %	6.531	3.350	90.1
Sangreal 15G-ICE	14.51 GB	0.134979	3.0024	13.4604	1.0226	93.10 %	6.531	3.350	89.6
≈ 12–13 GB
Sangreal 12G-ICEv2	11.90 GB	0.176176	3.9639	14.9417	1.0050	91.78 %	6.313	2.746	87.0
orcarouter Q2_K	12.94 GB	0.277227	6.7063	16.9488	1.0199	89.04 %	3.980	2.987	81.5

Table 3 — English text (WikiText-2)

PPL(base) 6.539079 · BF16 = 100

file	size	mean KLD	99% KLD	99.9% KLD	PPL ratio	same top-1	active bpw	file bpw	overall
≈ 28.5 GB
orcarouter Q6_K	28.51 GB	0.024367	0.1951	1.2990	0.9968	93.88 %	6.754	6.581	96.5
≈ 24.5 GB
Sangreal 25G-ICEv2	24.53 GB	0.024575	0.2211	1.2910	1.0013	93.74 %	7.385	5.662	96.4
Sangreal 25G-ICE	24.53 GB	0.025130	0.2162	1.1988	1.0016	93.64 %	7.385	5.662	96.4
≈ 22.5 GB
Sangreal 23G-ICEv2	22.52 GB	0.028207	0.2538	1.5501	1.0028	93.23 %	7.215	5.197	96.0
Sangreal 23G-ICE	22.52 GB	0.029232	0.2569	1.2277	1.0014	93.12 %	7.215	5.197	95.9
≈ 18–19 GB
Sangreal 19G-ICEv2	18.47 GB	0.043058	0.4310	1.6452	1.0078	91.59 %	6.872	4.263	94.6
Sangreal 19G-ICE	18.47 GB	0.050171	0.4730	2.1261	1.0079	90.92 %	6.872	4.263	93.9
orcarouter IQ4_XS	18.73 GB	0.058557	0.5249	2.2865	1.0056	89.86 %	5.076	4.323	93.1
≈ 16.8 GB
orcarouter Q3_K_M	16.76 GB	0.130658	1.1670	4.2389	1.0583	85.06 %	4.562	3.869	87.4
≈ 14.5 GB
Sangreal 15G-ICEv2	14.51 GB	0.095032	1.0010	3.6234	1.0412	87.56 %	6.531	3.350	90.2
Sangreal 15G-ICE	14.51 GB	0.101044	1.0494	3.3573	1.0524	87.08 %	6.531	3.350	89.7
≈ 12–13 GB
Sangreal 12G-ICEv2	11.90 GB	0.166550	1.7085	4.7462	1.0771	83.46 %	6.313	2.746	85.0
orcarouter Q2_K	12.94 GB	0.240005	2.0250	5.3516	1.1347	79.63 %	3.980	2.987	80.3

overall = 0.70/(1 + meanKLD) + 0.30 * sameTop1, ×100. BF16 = 100. The same composite as the Ornith-1.5 ICE card: 70% on how close the whole output distribution stays, 30% on agreement about the argmax.

Read the tail columns as shape, not order. 99.9% KLD is roughly the 33rd-worst token of 32,768 — an extreme order statistic with large sampling variance, so it inverts between adjacent files without meaning anything. Rank on mean KLD.

Two bpw columns. Only 8 of 256 experts fire per token, so a bit in ffn_*_exps is worth ~3% of a bit in attention, the shared expert or the output head. active bpw weights by that (token embedding excluded: it is a lookup, not a matmul); file bpw is just size ÷ parameters.

Don't compare the tables to each other. Code is more predictable text, so every file scores about half the divergence on it. Compare rows within a table.

Every row is measured against this trunk's own BF16 — our conversion of orcarouter's weights, published in this repo. Rivals are orcarouter's own quants of the same weights, so a difference here is the recipe and nothing else. These numbers are not comparable to the clean-trunk Nex repo's: a different reference.

What Sangreal is

The calibration corpus these tiers were quantized against. Built from primary sources, not assembled from eaddario's set, bartowski's calibration_datav5, or any other ready-made calibration file. 77 buckets, 9,533 documents, 47,443,549 tokens, sha256 85a6b823a6762bf6….

The full render is 92,663 chunks. These tiers use a 10,000-chunk pass over it, measured on Nex-N2.5-mini itself - 17x the 573-chunk imatrix behind the stock quants.

block	buckets	share	documents	synthetic
code + cyber + spec	36	45.10%	5,858	0.1%
reasoning	5	13.16%	553	100.0%
agentic	7	11.95%	1,018	65.2%
science + domain	11	11.24%	1,183	0.3%
language	11	9.66%	596	0.0%
mathematics	7	8.89%	325	0.0%
total	77	100%	9,533	21.7%

Split: technical 70.2 · science+domain 20.1 · language 9.7.

Synthetic is 21.7% by bytes and sits in exactly three buckets — teacher:k2-horizon, teacher:deepseek-v4-pro, teacher:hermes. Everything outside reasoning and agentic is real text.

Language — 11 languages, 5 non-Latin scripts

bucket	share	docs		bucket	share	docs
arabic	0.92%	48		russian	0.87%	58
chinese	0.89%	31		spanish	0.87%	72
hindi	0.88%	44		portuguese	0.87%	70
japanese	0.88%	33		turkish	0.87%	56
french	0.88%	60		german	0.86%	73
english	0.87%	51

The spread across the whole block is 0.06 points, and that is deliberate: language's job here is activation, not ranking. Each bucket sits just above the saturation floor at 44.9% consumption, so these buckets select for the first time — the legacy v1 spec had consumed 96% of turkish and 92% of english.

Mathematics — absent from the pool until this cycle

bucket	share	docs
math (general)	3.16%	170
analysis	1.24%	28
algebra	1.16%	46
geometry	1.12%	30
proof	1.08%	28
topology	1.05%	20
number-theory	0.06%	3

Sourced from the LaTeX algebraic geometry, mathlib4, UniMath, math-comp, open-web-math, AutoMathText and proof-pile.

The imatrix and the recipes are not published with this release.

What's inside

ICE = Isolation of Compounding Error: allocate bits by how far a quantization error travels, not by how large the activations are. Routers stay F32, SSM decay gates stay F32, the KV-cached projections stay F16, the always-on dense path stays Q8_0, and the budget goes to the routed expert bank — most of the parameters, but only 8-of-256 active per token. ICE 1.6 keeps token_embd at Q8_0 (its error propagates with context) and moves the output head to Q6_K (its error stays flat from 2K to 16K). Method and the cases where it does not win: gbuzhf/ICE-quantization.

No MTP head. The config declares one MTP layer, but the checkpoint carries no MTP weights. These tiers are the 40-block trunk; there is no built-in draft model. One consequence worth recording: a straight conversion copies mtp_num_hidden_layers into block_count, producing a GGUF that advertises 41 blocks and fails to load (check_tensor_dims: tensor 'blk.40.attn_norm.weight' not found). Every file in this repo declares 40, verified on each one.

Chat template: Qwen-Sharp v22.5.0-arrow4.0 (ours; spec below). Nex-N2.5-mini uses Qwen3.6's tokenizer unchanged (vocabulary, merges and special tokens; seven extra audio/TTS tokens), so the template's delimiters are native tokens — checked on every tier at signing.

Vision. Abliteration edits the vision merger's output projection, so this repo ships a projector built from orcarouter's own weights, converted to Q8_0 (334 tensors). Load it with --mmproj mmproj-Nex-N2.5-mini-OrcaRouter-Q8_0.gguf.

Chat template — Qwen-Sharp v22.5.0-arrow4.0

Every GGUF embeds arrow4.0. To pin it explicitly, pass the file shipped here:

llama-server -m <model>.gguf --jinja --chat-template-file chat_template_sharp-v22.5.0-arrow4.0-nex.jinja

arrow4.0 is our agent layer on Qwen-Sharp v22.5.0. Plain chat stays byte-identical to Sharp; with tools it adds short agent rules, one note per tool result, a STATUS line, thinking only at decision points, .arrow project memory and a hand-back report when work stalls. The -nex file opens replies the Nex way, so the harness's reasoning level is respected (default medium = adaptive).

Full description and switches: gbuzhf/Qwen-Sharp-Arrow-Chat-Templates.

Files

file	size	sha256
Nex-N2.5-mini-OrcaRouter-Sangreal-12G-ICEv2.gguf	11.90 GB	c93241e5bac5…
Nex-N2.5-mini-OrcaRouter-Sangreal-15G-ICEv2.gguf	14.51 GB	3553ac0e0ff4…
Nex-N2.5-mini-OrcaRouter-Sangreal-19G-ICEv2.gguf	18.47 GB	788a0cd9f818…
Nex-N2.5-mini-OrcaRouter-Sangreal-23G-ICEv2.gguf	22.52 GB	5acca1d6753e…
Nex-N2.5-mini-OrcaRouter-Sangreal-25G-ICEv2.gguf	24.53 GB	2945ec2c74da…
Nex-N2.5-mini-OrcaRouter-Sangreal-15G-ICE.gguf	14.51 GB	5d0f82d7883a…
Nex-N2.5-mini-OrcaRouter-Sangreal-19G-ICE.gguf	18.47 GB	746c807b268d…
Nex-N2.5-mini-OrcaRouter-Sangreal-23G-ICE.gguf	22.52 GB	2537692ba0d7…
Nex-N2.5-mini-OrcaRouter-Sangreal-25G-ICE.gguf	24.53 GB	36027951b141…
mmproj-Nex-N2.5-mini-OrcaRouter-Q8_0.gguf	0.61 GB	dab66e630f44…
Nex-N2.5-mini-OrcaRouter-Q8_0.gguf	36.90 GB	df18ea001106…
BF16/Nex-N2.5-mini-OrcaRouter-BF16-00001-of-00002.gguf	44.87 GB	66b630aa5cf2…
BF16/Nex-N2.5-mini-OrcaRouter-BF16-00002-of-00002.gguf	24.51 GB	aebfe205394f…

• measurements/2026-10-10/ — raw llama-perplexity output for every ICEv2 row and for every ICE 1.6 and orcarouter file re-measured in the same session (× 3 columns), plus the three BF16 base runs.
• measurements/2026-10-04/ — raw llama-perplexity output for every row of every table (4 ICE tiers + 4 orcarouter files × 3 columns), the three BF16 base runs, and the session summary.

Credits

base model	nex-agi/Nex-N2.5-mini
base model	orcarouter/Nex-N2.5-mini-Uncensored
comparison ladder	orcarouter/Nex-N2.5-mini-Uncensored-GGUF
chat template	peculiar-ragdoll/Qwen-Sharp-Chat-Templates
ICE method	gbuzhf/ICE-quantization

KLD measures fidelity to this repo's reference and nothing else — not reasoning, tool use or speed. It is also within-model: every file here is measured against this trunk's own BF16, so these values are not comparable to another repo's. Compare across models with PPL.