PollardWeights/LFM3-Audio-2B-Pollard

LFM3-Audio-2B -- Pollard

Pollard shrank this model: 3.96 GB (f16) -> 0.83 GB -- 79% smaller, 4.8x down.

The smallest rung here; larger, higher-fidelity rungs are listed below.

format	this model's size
f16	3.96 GB
Q8_0	~2.10 GB
Q6_K	1.43 GB
Q4_K_M	~1.15 GB
PollardMix (this repo's IQ3_S)	0.83 GB

Pollard builds of LiquidAI/LFM3-Audio-2B made with Pollard Weights -- a ladder of measured-allocation quants (bits placed by per-layer sensitivity, not a uniform crush).

Standard GGUF -- every file here runs in stock llama.cpp / ik_llama.cpp, Ollama, LM Studio.

Model details

Parameter count	~2.0B
Architecture	lfm3_audio
Input support	text, audio
imatrix	yes -- see calibration
Perplexity measured	yes -- table below

Which file should I choose?

Every rung is the same weights, sized to a different RAM budget by the measured allocation. Pick the largest one that fits your machine with room for context:

• ~3.4 GB RAM / VRAM -> Q6_K (1.43 GB).
• ~3.1 GB RAM / VRAM -> IQ4_XS (1.07 GB).
• ~2.8 GB RAM / VRAM -> IQ3_S (0.83 GB).

Available files

file	PPL	size	Mean KLD	Top-1 agree	Tool-ready	verified	notes
LFM3-Audio-2B-Pollard-IQ3_S.gguf	3.5624	0.83 GB	0.2271	83.09%	yes (106% of f16; CI 96-117)	yes	smallest
LFM3-Audio-2B-Pollard-IQ4_XS.gguf	3.2028	1.07 GB	0.0347	93.01%	yes (100% of f16; CI 94-106)	yes
LFM3-Audio-2B-Pollard-Q6_K.gguf	3.1532	1.43 GB	0.0036	97.45%	yes (100% of f16; CI 100-100)	yes	recommended default

Top-1 agree = share of tokens where the rung's most-likely token is the same one the f16 would have picked (llama-perplexity's Same top p for GGUF, the same logit comparison in pollard-verify for the other lanes). Higher is closer to the original model.

Verified = the rung passed pollard-verify --gate: it is the complete model (every tower, projector and processor file of the source), it loads, and it passes the coherence and runtime checks. Its KLD / top-1, tool-call and retention numbers are measured and shown here as they are; "below band" names a number outside the gate file's band for its bit width -- reported, not blocking.

Tool-ready = the rung still calls tools (pollard-toolcall): served with llama-server --jinja, asked 60 tool probes (single call, choose among eight tools, no call needed, nested arguments) through OpenAI-style tools, and scored on the calls that server parses -- shown as success relative to the reference (the f16 for GGUF, the bf16 original for the other lanes) with a 95% paired-bootstrap CI, or as a bare success rate when no reference run was recorded. "cross-backend" = the reference ran on a different server than the rung; "not measured" = this lane's server could not run where the rung was built; "error" = request failures prevented a valid capability measurement.

Retention vs f16

What each rung keeps of the f16 on real work -- quantized score / f16 score on the same tasks, the same settings, the same harness (100% = no loss). Measured against our own f16, so the rungs compare directly; it is not comparable to a retention figure published on a different suite.

file	Tool calls
LFM3-Audio-2B-Pollard-Q6_K.gguf	100.0%
LFM3-Audio-2B-Pollard-IQ4_XS.gguf	100.0%
LFM3-Audio-2B-Pollard-IQ3_S.gguf	105.8%

Tool calls = the rung's tool-call success as a share of the f16's (pollard-toolcall); "valid" when no f16 run was recorded. Decisions = share of typed decisions identical to the f16's.

Download a specific file

pip install -U "huggingface_hub[cli]"
hf download PollardWeights/LFM3-Audio-2B-Pollard \
  --include "LFM3-Audio-2B-Pollard-IQ3_S.gguf" --local-dir ./

How to run

These are standard GGUF and run with llama.cpp:

llama-server -hf PollardWeights/LFM3-Audio-2B-Pollard:IQ3_S

or from a local file:

llama-cli    -m LFM3-Audio-2B-Pollard-IQ3_S.gguf -ngl 99 -p "Explain why the sky is blue."
llama-server -m LFM3-Audio-2B-Pollard-IQ3_S.gguf -ngl 99      # OpenAI-compatible API + web UI at :8080

They also work in anything built on llama.cpp -- LM Studio, koboldcpp, Jan, ramalama, Ollama (ollama run hf.co/PollardWeights/LFM3-Audio-2B-Pollard).

Tool calling

The chat template writes tool calls in the qwen3_coder syntax (detected by rendering it with pollard-modelkind):

<tool_call>
<function=get_weather>
<parameter=city>
Paris
</parameter>
<parameter=days>
7
</parameter>
</function>
</tool_call>

Serve with:

• llama.cpp (GGUF): llama-server -m LFM3-Audio-2B-Pollard-IQ3_S.gguf --jinja -- the model's own template renders the tools and llama.cpp parses the calls (specialised: Qwen3-Coder); pass tools in /v1/chat/completions.
• vLLM: vllm serve <the original weights> --enable-auto-tool-choice --tool-call-parser qwen3_coder

Use it in your harness

Any agent harness that speaks the OpenAI, Anthropic or Responses API can drive this model through llama-server. Pollard Weights' pollard serve picks the runtime this rung needs, sizes the context to the harness and prints its exact config (--write merges it in, after backing the original up):

pollard serve PollardWeights/LFM3-Audio-2B-Pollard:IQ3_S --harness claude-code      # or omp, codex, opencode, aider, ...

harness	calls	min. context	setup
Oh My Pi (omp)	/v1/chat/completions	32k	~/.omp/agent/models.yml: provider api: openai-completions, baseUrl: http://127.0.0.1:8080/v1, auth: none
Claude Code	/v1/messages	64k	ANTHROPIC_BASE_URL=http://127.0.0.1:8080 ANTHROPIC_AUTH_TOKEN=pollard ANTHROPIC_MODEL=LFM3-Audio-2B-Pollard-IQ3_S
Codex CLI	/v1/responses	32k	~/.codex/config.toml: [model_providers.pollard] base_url = "http://127.0.0.1:8080/v1", wire_api = "responses"
OpenCode	/v1/chat/completions	32k	opencode.json: provider npm: @ai-sdk/openai-compatible, options.baseURL: http://127.0.0.1:8080/v1, limit.context
Cline / Roo Code	/v1/chat/completions	32k	API Provider "OpenAI Compatible", Base URL http://127.0.0.1:8080/v1
Aider	/v1/chat/completions	16k	OPENAI_API_BASE=http://127.0.0.1:8080/v1 aider --model openai/LFM3-Audio-2B-Pollard-IQ3_S
Continue	/v1/chat/completions	32k	config.yaml: provider: llama.cpp, apiBase: http://127.0.0.1:8080
Goose	/v1/chat/completions	32k	GOOSE_PROVIDER=openai OPENAI_HOST=http://127.0.0.1:8080
OpenHands	/v1/chat/completions	32k	model openai/LFM3-Audio-2B-Pollard-IQ3_S, base URL http://host.docker.internal:8080/v1
Open WebUI	/v1/chat/completions	8k	Admin Settings -> Connections -> OpenAI API: http://127.0.0.1:8080/v1
Ollama	own runtime	8k	ollama run hf.co/PollardWeights/LFM3-Audio-2B-Pollard:IQ3_S, then /set parameter num_ctx 8192 or more
LM Studio / Jan	own runtime	8k	load the GGUF, set the context length to 8192 or more

Without Pollard, start llama-server yourself with a context the harness fits in, the model's own template (tool calls are parsed through it) and the thinking kept out of the answer:

llama-server -hf PollardWeights/LFM3-Audio-2B-Pollard:IQ3_S -c 65536 --jinja --reasoning-format deepseek --alias LFM3-Audio-2B-Pollard-IQ3_S -np 1 --port 8080

Too small a context fails silently: the harness's system prompt and tool schemas are truncated and the model looks like it ignores its tools. Claude Code needs 64k.

imatrix (calibration)

The importance matrix (LFM3-Audio-2B.imatrix, included) was computed on the pollard-calib 'auto' profile (sources, revisions and licences in the shipped .sources.json).

The exact corpus is included as release-calib.txt, so the allocation can be reproduced rather than taken on trust.

ARM / AVX

llama.cpp repacks weights into an interleaved layout at load time for faster inference on ARM and AVX machines -- no special file needed, online repacking covers these quants. The old Q4_0_4_4/4_8/8_8 variants are not required.

Errata

• Every file here loads in stock llama.cpp -- verified from the tensor types with pollard-ggufcheck, not assumed from the filenames.
• Measured allocation places bits by per-layer sensitivity under a size budget.
• Single machine; replication invited.

Credits & license

• Base model: LiquidAI/LFM3-Audio-2B (LiquidAI)
• Quantization tooling: llama.cpp (ggml-org)
• Method + tooling: Pollard Weights -- measure first, no claim before a number.
• License: other, inherited from the base model.

Built with Pollard Weights -- frontier models, small hardware, no compromise.