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):

† 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).

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

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.

Table 3 — English text (WikiText-2)

PPL(base) 6.539079 · BF16 = 100

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.

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

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

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

• 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

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.