Runtime: Osaurus 0.25.32 or newer.

OsaurusAI/Ripple-65B-Preview-JANG_4D

Ripple-65B Preview in 43.7 GB, with 98.3% top-1 agreement with the full-precision model. The model's DFlash2 speculative-decoding drafter is included.

This is a JANG bundle of ricursiveAI/Ripple-65B-Preview: a 65B dense transformer with 80 layers. Every fifth layer is full attention without positional rotation and the others are sliding-window attention (1,024 tokens), with gated attention and a value residual.

• Weights: affine 4 to 8 bits (group 64), allocated per module from measured sensitivity at the JANG_4M size budget, with attention and the output head kept at 6 bits or more.
• Rounding: GPTQ on every attention projection, gate/up and lm_head; activation-weighted rounding (imatrix) on down_proj, where it measured better; embeddings 6-bit.
• Activation-aware scaling (AWQ) folded exactly into the feed-forward input norm.
• Size: 43.67 GB of model shards, 5.39 bits per weight over the 64.8B source parameters.
• Drafter (dflash2/): ricursiveAI/Ripple-65B-Preview-DFlash2, unmodified BF16 (8.08 GB). Osaurus quantizes it to 4-bit when it loads.

Smaller sibling: OsaurusAI/Ripple-65B-Preview-JANGH2 (24.6 GB).

Quality vs the original model

The test set is 221,809 assistant positions, 30,000 or more in each of 7 domains, scored against the BF16 release run on the original code, with top-128 KL. The assistant turns were written by Ripple itself. None of these conversations were used for calibration; the choice between rounding methods per module did look at their input rows, so the scores may be slightly optimistic.

	mean KL ↓	p99 KL ↓	top-1 ↑
BF16 on this runtime (measurement floor)	0.00017	0.0016	99.48%
Ripple-65B-Preview-JANG_4D (43.7 GB)	0.00239	0.0263	98.30%
same size, fixed tiers (JANG_4M)	0.00325	0.0390	98.09%
Ripple-65B-Preview-JANGH2 (24.6 GB)	0.0841	0.939	90.28%

Change in negative log-likelihood of the held-out tokens: +0.002.

By domain (mean KL · top-1):

domain	positions	JANG_4D	JANGH2
agentic tool loops	30,839	0.0018 · 98.6%	0.066 · 91.3%
coding	36,021	0.0026 · 97.9%	0.081 · 89.8%
terminal	32,049	0.0017 · 98.6%	0.057 · 91.3%
reasoning / math	31,341	0.0018 · 98.5%	0.060 · 91.1%
cybersecurity	31,076	0.0032 · 98.0%	0.122 · 88.3%
multilingual, science, creative	30,457	0.0041 · 97.7%	0.153 · 86.9%
Osaurus app tasks	30,026	0.0016 · 98.8%	0.052 · 93.3%

Tool decisions and reasoning

• Starting a tool call: at 250 points where the original model starts a tool call, JANG_4D's top choice matches at 248 (99.2%).
• Closing reasoning: at all 256 points where the transcript closes the model's reasoning (</think>), JANG_4D's top choice is </think>.
• First answer token: matches at 99.2% of 364 turns; end of turn at 99.7% of 336.

Speculative decoding

dflash2/ holds the model's own DFlash2 drafter, which proposes up to 15 tokens per step for the main model to verify. Every drafted token is verified by the main model, so output follows the main model's distribution. Osaurus picks the draft length on the fly. On the capture GPU with the full-precision model, the drafter's first token was accepted 68-83% of the time and it got 2.0-5.0 drafted tokens accepted per step, depending on the domain (code and agentic work about 3.3-3.4). Speed on a Mac depends on the machine and is not stated here.

Serving contract

• Reasoning: the model decides on each turn whether to think. The chat template ends at the assistant header and never opens <think>. There is no thinking switch and there are no effort levels. Earlier reasoning stays in the history by default (clear_thinking=false).
• Tool calls: JSON inside <tool_call> ({"name": ..., "arguments": {...}}); tool results go back as <tool_response> blocks in a user turn. For tool_choice: "required", use strict tools ("strict": true, additionalProperties: false), as the original model card recommends.
• Sampling: temperature 1.0, top-p 0.95 (generation_config.json), the model's distribution unchanged. The original release ships no sampling defaults.
• EOS: <|im_end|>, <|endoftext|>. No BOS token in chat.
• Context: the original card states up to 1M tokens. Only the 16 full-attention layers keep a KV cache that grows with the context.

What's in the bundle

All bit widths below were measured from the shard bytes.

• q / k / attention-gate projections: 6-bit (×80 each).
• v_proj: 8-bit ×57, 6-bit ×23. o_proj: 6-bit ×74, 8-bit ×6.
• gate_proj: 4-bit ×43, 5-bit ×35, 6-bit ×2. up_proj: 4-bit ×40, 5-bit ×36, 6-bit ×4. down_proj: 4-bit ×46, 5-bit ×24, 6-bit ×10.
• lm_head 8-bit, embeddings 6-bit. Norms and per-layer value-residual weights bf16.
• dflash2/: the drafter (model.safetensors, config.json, its README.md, LICENSE, provenance.json with the source revision and sha256).

Build details

• Calibration: 1.52M tokens of Ripple's own outputs and real tool loops. The mix is coding across many languages, multi-turn agentic tool loops, reasoning/math, terminal sessions, cybersecurity, and multilingual/science/creative text. Long rows up to 30k tokens, so the full-attention layers see real long context. Held-out conversations were split off first by source.
• Hessians: per-linear input statistics from a layer-streamed run of the BF16 model in its original code.
• Allocation: each module's error at every width was converted to a predicted KL cost with per-class coefficients measured end to end, then bits were filled under the size budget. This beat the fixed-tier JANG_4M at the same size on held-out KL (0.00239 vs 0.00325) and top-1 (98.30% vs 98.09%).
• Method per module: GPTQ or imatrix, chosen on held-out input rows; GPTQ won on 561 of 641 choices. Against an imatrix-only build of the same allocation, it was better in all 7 domains (KL 0.00239 vs 0.00333).
• Verification: every tensor file is aligned, there is one index, every module's bits match the config, and every source tensor is accounted for. The bundle loaded through the runtime scores within 0.00001 KL of the offline measurement.