ISTA IQ3_S vs HauhauCS IQ3_XS @16 GB: features
Both 3-bit Qwen3.8-27B builds pass vision, tool calling (with and without thinking), using tool results and thinking control on my RTX 4070 Ti SUPER, at ~43–44 tok/s with 64K fully on the GPU.
- Tested
- Setup
GeForce RTX 4070 Ti SUPER 16 GB — Proxmox VM (GPU passthrough)Proxmox VE guest with PCIe passthrough, Ubuntu 24.04, NVIDIA 595 open driver · Intel Core i5-13400 host; guest 8 vCPU · Guest 32 GB; host 64 GB DDR5
- Runtime
- Ollama 0.34.1, q8_0 KV cache
- Models
Qwen3.8-27B GSQ-RCOISTA IQ3_Stested
Qwen3.8-27B Aggressive MTPIQ3_XStested
Verdict
A tie on features: every functional check passes on both, and speed is within 1.5%. Choosing between them has to come down to answer quality on real work — see the CVE evaluation that followed.
Works
- 6/6 functional checks pass on both: vision, tool call, tool result, thinking on/off, tools + thinking
- 65,536 tokens fit 100% on the GPU with q8_0 KV
- Warm vision answers in 0.2–0.7 s
Doesn’t
- ISTA's first (cold) vision request took 76.6 s while the projector loaded
- Above 64K both spill to the CPU (81,920 → ~91% GPU)
- Decode, ISTA / HauhauCS
- 43.8 / 43.3tok/s
- Functional checks
- 6/6 · 6/6
- Largest all-GPU context
- 65,536tokens
VRAM vs context (q8_0 KV)
Model size reported by ollama ps
- ISTA IQ3_S
- HauhauCS IQ3_XS
Data
| ISTA IQ3_S | HauhauCS IQ3_XS | |
|---|---|---|
| 16,384 | 11.65 GiB | 11.70 GiB |
| 32,768 | 12.26 GiB | 12.31 GiB |
| 49,152 | 12.87 GiB | 12.92 GiB |
| 65,536 | 13.48 GiB | 13.53 GiB |
| 81,920 (partial offload) | 14.77 GiB | 14.97 GiB |
| 98,304 (partial offload) | 15.53 GiB | 15.63 GiB |
Why
Two 3-bit builds of Qwen3.8-27B fit on a 16 GB card: ISTA-DASLab’s GSQ-RCO IQ3_S (from a quantization research group, with a measured perplexity lead) and HauhauCS “Aggressive” IQ3_XS (an uncensored fine-tune with a built-in MTP head). Before testing quality at all, I wanted to know: does everything I rely on actually work through Ollama?
Functional checks
Each check is a plain pass/fail against a fixed expectation, and I wrote the prompts in Turkish.
| Check | What it asks | ISTA IQ3_S | HauhauCS IQ3_XS |
|---|---|---|---|
| Vision | Read the text and shapes in a generated image (“KAPI KODU: 4821”, red circle left, blue square right) | ✓ | ✓ |
| Tool call | Call get_weather(city, unit) for İzmir in Celsius, no text leaking into content |
✓ | ✓ |
| Tool result used | Answer from the returned tool result (27.4 °C, clear) | ✓ | ✓ |
| Thinking on | Answer 231 with a separate thinking field | ✓ (189 chars) | ✓ (219 chars) |
| Thinking off | Answer 231 with no thinking | ✓ (4 tokens) | ✓ (4 tokens) |
| Tool call with thinking | Same tool call while thinking is on | ✓ | ✓ |
Thinking levels: low → 188 / 196 thinking characters, high → 307 / 238 — a small, inconsistent effect on a trivial
prompt.
Speed
| ISTA IQ3_S | HauhauCS IQ3_XS | |
|---|---|---|
| Decode (3 runs) | 43.9 / 43.8 / 43.7 tok/s | 43.4 / 43.3 / 43.3 tok/s |
| Vision, cold (projector load) | 76.6 s | 4.8 s |
| Vision, warm | 0.6 s, 0.2 s | 0.7 s, 0.2 s |
Context fit (q8_0 KV)
| num_ctx | ISTA size | GPU | HauhauCS size | GPU |
|---|---|---|---|---|
| 16,384 | 11.65 GiB | 100% | 11.70 GiB | 100% |
| 32,768 | 12.26 GiB | 100% | 12.31 GiB | 100% |
| 49,152 | 12.87 GiB | 100% | 12.92 GiB | 100% |
| 65,536 | 13.48 GiB | 100% | 13.53 GiB | 100% |
| 81,920 | 14.77 GiB | 91.7% | 14.97 GiB | 90.4% |
| 98,304 | 15.53 GiB | 86.7% | 15.63 GiB | 85.1% |
With q8_0 KV, the practical all-GPU limit is 65,536 tokens for both (vs 49K with fp16 KV).
A first look at real work
As a quick smoke test, I had both models triage Log4Shell (CVE-2021-44228) through my chat UI with the CVE tools. Both got it right (risk 100/CRITICAL, CVSS 10.0, EPSS 0.99999, listed in CISA KEV, upgrade path 2.16.0 → 2.17.0 → 2.17.2) in ~58 s at ~41–42 tok/s. That was too easy to tell them apart — hence the full CVE evaluation.