SNEX-V1 (Sage-Nex) — GGUF

SNEX-V1 is Sage-A1.5 with a small, sparse transplant of Nex-N2.5-mini's training. Sage is the body; from Nex only the largest changes its own training made are carried over, at a quarter of their strength. Everything else is Sage's: the vision projector, the MTP draft head, the router, the norms, the embeddings.

An experimental merge: it is not claimed or shown to improve on Sage-A1.5.

Fidelity of the ICE 2.0 tiers (vs the BF16 master)

Score: BF16 = 100; 70 % full-distribution fidelity 1/(1+KLD) + 30 % same-top-1, averaged over the three sets.

How SNEX was built

1. Lineage first. Weight comparisons placed the three models: Sage-A1.5 is a fine-tune of Qwen3.6-35B-A3B; Nex-N2.5-mini is a fine-tune of Qwen3.5-35B-A3B; Qwen3.6 continues Qwen3.5. So Nex's own training is τ = Nex − Qwen3.5. Taking the difference against Qwen3.6 instead would also subtract the 3.5 → 3.6 upgrade that Sage already carries.

2. What is transplanted. SNEX = Sage + 0.25 · TopK₂₅%(τ), done directly on the GGUF tensors:

• 430 weight matrices take part (attention, DeltaNet projections, shared expert, every routed expert). For each one, only the 25% of entries where Nex's training moved the weights most are kept; for the routed experts the cut is made per expert, so every expert keeps its own top quarter. The kept 25% hold ~73% of the delta's energy.
• The kept delta is added at weight 0.25. Nex's training direction is close to orthogonal to Sage's own (cosine ≈ 0 on the routed experts), so the two add on top of each other instead of fighting.
• 303 tensors are frozen to Sage exactly: token embeddings, output head, every norm, the router and shared-expert gates, and the DeltaNet state parameters (A, dt bias, conv). The routing decisions, the vocabulary interface and the recurrent dynamics stay Sage's.
• RMSNorm weights are handled in GGUF's stored form (+1 folded in) for all three models before any arithmetic.

3. Sage's extras, untouched. Sage's MTP draft head (blk.40) is removed before the arithmetic and grafted back afterwards, so speculative decoding keeps Sage's trained draft layer. Sage's vision projector is shipped as is.

4. Chat template: Arrow 4.0-snex. Arrow 4.0 with an adaptive think opener: at reasoning_effort low / medium / high the assistant turn opens with <think> and the model decides how long to think; xhigh opens with <think>\n. Pass adaptive_thinking: false in chat_template_kwargs to get the plain Arrow 4.0 opener at every effort.

5. Quantization: Sangreal + ICE.

• Importance matrix: the Sangreal calibration corpus (code in 29 languages, 10 natural languages, agentic traces, science, medicine, security…), 10,000 chunks of 512 tokens, measured on SNEX-V1's own Q8_0. Never borrowed from another model.
• ICE 1.6: per-tensor quantization recipes (attention and shared paths kept wide, routed experts sized per tier, the MTP block kept at draft-safe precision).
• ICE 2.0: after quantization, the routed experts are repaired: their scales (and for the IQ tiers their codes) are refit against the real activations of the full-precision model on Sangreal text, so the quantized experts reproduce the BF16 outputs more closely than rounding alone.

Use

llama-server -m SNEX-V1-19G-ICEv2.gguf --mmproj mmproj-SNEX-V1-F16.gguf --jinja -c 65536 -fa on

The MTP head is in every file; on a llama.cpp build with MTP support, enable it for speculative decoding (draft depth 1).

Files

Credits

Sage-A1.5 by InternScience, Nex-N2.5-mini by nex-agi, Qwen3.5 / Qwen3.6 by the Qwen team, all Apache-2.0. Merge, Sangreal corpus, ICE recipes and quantization by gbuzhf.