Add files using upload-large-folder tool

README.md

@@ -22,24 +22,10 @@ This is a local, self-quantized GGUF build of [XiaomiMiMo/MiMo-V2.5](https://hug
 
 This quant was optimized for systems with 128 GB of memory. The default serving profile targets a 128 GB Apple Silicon machine and tries to keep the model practical at a 100,000-token context. Smaller-memory systems will likely need more aggressive CPU offload, a smaller context, or a different quant.
 
-It is a text-only llama.cpp conversion of the MiMo language backbone. The original MiMo-V2.5 checkpoint is omnimodal, but this GGUF does not include the vision or audio encoders. The MiMo multi-token prediction blocks were also omitted during conversion because normal llama.cpp generation does not currently execute them for this model.
-
-## Files
-
-The model is split into 16 GGUF shards:
-
-```text
-MiMo-V2.5-coder-Q2-00001-of-00016.gguf
-...
-MiMo-V2.5-coder-Q2-00016-of-00016.gguf
-```
-
-Load the first shard. llama.cpp will find the remaining shards automatically.
+It is a text-only quantization. The original MiMo-V2.5 checkpoint is omnimodal, but this GGUF does not include the vision or audio encoders. The MiMo multi-token prediction blocks were also omitted during conversion because normal llama.cpp generation does not currently execute them for this model.
 
 ## Quantization
 
-This artifact was quantized from the original XiaomiMiMo checkpoint, not from a third-party GGUF.
-
 High-level summary:
 
 - Quant type: `Q2_K_S`
@@ -50,17 +36,9 @@ High-level summary:
 
 One tokenizer metadata fix is included so llama.cpp does not warn about the base-vocab `</s>` token at load time. MiMo's actual EOS token remains `<|im_end|>`.
 
-## Importance Matrix
-
-The importance matrix is what makes this quant more targeted than a generic low-bit conversion.
-
-It was built from English coding and agent-style prompts: reading files, searching code, running shell commands, editing workflows, short code review tasks, and OpenAI-compatible tool calls. The tool-call calibration included realistic argument shapes such as bounded file reads, tail reads, grep context lines, and command arrays.
-
-That means the quantization tries to spend its limited precision budget on weights that matter for coding and structured tool use. It is not a general-purpose multilingual calibration set, and it was not designed to preserve Chinese or multimodal quality.
+## Imatrix
 
-## Why This Quant Exists
-
-The goal is not to preserve every capability of the original model equally. This build deliberately prioritizes:
+This build deliberately prioritizes:
 
 - reliable OpenAI-compatible tool calls
 - coding and shell-oriented agent use
@@ -71,8 +49,6 @@ Chinese-language quality and multimodal use were not optimization targets.
 
 ## Serving
 
-Install or build llama.cpp with `llama-server` available in `PATH`. Once the model is on Hugging Face, the usual way to run it is directly from the Hub:
-
 ```sh
 llama-server \
   -hf jedisct1/MiMo-V2.5-coder-Q2 \
@@ -193,32 +169,13 @@ llama-server \
 
 For best tool-calling results:
 
-- Use OpenAI-compatible request-provided tool schemas.
-- Keep llama.cpp built-in tools disabled unless you are specifically testing them.
+- Use the [Swival](https://swival.dev) harness
 - Disable model reasoning output with `--reasoning off` or `MIMO_REASONING=off`.
 - Set `parallel_tool_calls` to `false` if your client supports it.
 - Avoid forcing `tool_choice: required`; in testing it made malformed calls more likely.
 
 This build was tested with Swival-style tool schemas for `read_file`, `grep`, `outline`, `run_command`, and `todo`.
 
-## Local Test Results
-
-On the local 128 GB Apple Silicon M5 setup:
-
-- The first local `Q2_K` artifact passed only 5/7 in its best Swival-shaped serving configuration.
-- This imatrix-backed `Q2_K_S` artifact passed 21/21 across three Swival-shaped repeat runs with fast M5 serving defaults.
-- A Swival-style `Hello` request rendered to 4,411 prompt tokens because the client included a system prompt and tool schemas. With fast M5 serving, llama.cpp processed that payload at about 239 prompt tokens/sec. With all MoE tensors on CPU, the same class of prompt processed at about 13 prompt tokens/sec.
-
-These are local smoke and agent-harness results, not a public benchmark suite.
-
-## Limitations
-
-- Text-only GGUF: no vision, video, or audio encoders.
-- MTP blocks are omitted.
-- The quantization is very low bit. It is intended to fit and run locally, not to match the full BF16 checkpoint.
-- The default 100,000-token context is much smaller than MiMo-V2.5's advertised 1M training context, but much more practical on this hardware.
-- Quality should be validated on your own coding and tool-calling workloads before relying on it.
-
 ## License
 
 The upstream model card for `XiaomiMiMo/MiMo-V2.5` declares the MIT license. This derived GGUF is provided under the same license metadata.